Research · The Agent Harness Landscape and the Shift to Markdown-Defined, Event-Driven Agents
Back to researchResearch sweep · deep · 2025 – 2026
The Agent Harness Landscape and the Shift to Markdown-Defined, Event-Driven Agents
How coding and general-purpose agent harnesses evolved between October 2025 and October 2026 (Claude Code, OpenAI Codex CLI, OpenCode, Pi, Antigravity CLI, Cursor, Amp, OpenClaw): adoption and user base, the move to event-driven and multi-agent workflows defined in markdown, the SKILL.md Agent Skills standard and its most popular skills repositories, and whether harness scaffolding stays necessary as models improve
Synthesised 2026-10-25
Harnesses, 2025 to 2026: the scaffolding moved, it did not shrink
Overview
In twelve months the coding agent harness went from a wrapper around a chat model to the product itself. Anthropic reported Claude Code run-rate revenue above $2.5bn in February 2026, doubled since 1 January, and OpenAI counted more than 5 million weekly Codex users by June. Gartner published its first Magic Quadrant for the category in May and sized it at roughly $10bn annualised. Those are three independent measures of the same thing: the loop, the permissions, the triggers and the files that steer a model now carry commercial weight on their own. Sources: Constellation Research (2026) (↗); Constellation Research (2026) (↗); Gartner (2026) (↗)
The landscape sorts into three shapes. Closed vendor harnesses tied to one lab (Claude Code, Codex CLI, and from May 2026 Google's Antigravity CLI, a closed Go binary that replaced the open Gemini CLI). Open, model-agnostic, batteries-included harnesses (OpenCode, OpenClaw, Amp). And minimal-core harnesses, led by Mario Zechner's Pi, which shipped without MCP, subagents or plan mode and only added MCP to its core on 29 September 2026. Cursor sits across categories as an IDE harness that now belongs to SpaceX after an all-stock deal closed on 14 August 2026. Sources: Virtualization Review (2026) (↗); Implicator.ai (2026) (↗); explainx.ai (2026) (↗); Wikipedia (2026) (↗
The defining shift was from chat to delegation. The October 2026 Claude Code changelog adds scheduled and unattended retry modes, Slack-triggered sessions, cloud sessions and Managed Agents subagent teams. OpenCode 2.0 repositioned itself as a background service with an API and SDK. Google justified killing Gemini CLI by saying developer work had moved to multiple agents splitting tasks over a shared backend. The developer loop is now a set of events that wake agents, not a terminal someone types into. Sources: Anthropic (Claude Code Docs) (2026) (↗); Shattered (2026) (↗); The New Stack (2026) (↗)
The second shift was that plain markdown became the control plane. AGENTS.md, CLAUDE.md and SKILL.md survived every rewrite in the period, including Google's switch from open to closed source. The question this sweep tries to settle is whether that layer is durable or a stopgap the models will absorb. The answer, argued below, is that the format is durable and most of its content is not.
Timeline
- Anthropic launches Agent Skills and the SKILL.md format
- Claude Code passes $1bn run rate, Cursor raises at $29.3bn
- Codex adopts skills and the open spec is published
- Claude Code run rate passes $2.5bn
- Skills registry listings jump from 2,179 to 40,000 in three weeks
- ETH Zurich finds AGENTS.md gives no task-success gain
- OpenClaw founder joins OpenAI, OpenClaw moves to Codex auth
- Anthropic cuts third-party harnesses from Pro and Max plans
- Pi acquired by Earendil, Thoughtworks puts coding agents on a leash
- Google retires Gemini CLI for closed Antigravity CLI
- Gartner publishes first coding agent Magic Quadrant
- Codex passes 5m weekly users
- SpaceX completes Cursor acquisition
- Academic source-code studies of eleven harnesses appear
- MCP enters Pi core, scaffold effect named as evaluation variable
- Pi 1.0 ships, Stack Overflow survey puts Claude Code at 65.5%
- Claude Code, Codex and OpenCode 2.0 ship within four days
- Anthropic donates Agent Skills to the Linux Foundation AAIF
Key Findings
The hard numbers all rank Claude Code first, but they measure different things. Stack Overflow's 2026 Developer Survey (12,255 respondents, published 6 October) found 65.5% had used Claude Code in the past year, 58.7% Copilot, 29.5% Codex, 25.5% Cursor, 16.0% Antigravity, 13.6% OpenCode, 4.4% Pi and 2.1% Aider. That is the only like-for-like series across harnesses. Revenue (Claude Code), weekly users (Codex), stars (OpenCode at 214k against 151k for Claude Code) and valuations (Cursor at $60bn) are not comparable with one another, and the lanes that ranked them together did so loosely. Sources: Stack Overflow (2026) (↗); Morph (2026) (↗); Bloomberg (2026) (↗)
Lab pricing shaped harness choice more than features did. From 4 April 2026 Anthropic stopped Pro and Max limits covering OpenClaw and other third-party harnesses, requiring metered "extra usage". OpenClaw's creator Peter Steinberger had already joined OpenAI in February and moved the project to Codex OAuth. The financial-press and practitioner lanes agree on the ban; neither retrieved a primary source on a reversal, so the "reversal" in the brief is untraced in this sweep. Sources: The Next Web (2026) (↗); TechBuzz (2026) (↗); Hacker News (2026) (↗)
Markdown directive files survived a rewrite, and then failed a controlled test. Antigravity CLI reads AGENTS.md and keeps skills, hooks and subagents despite a new codebase and closed licence. Yet Gloaguen et al. at ETH Zurich found repository context files produced no improvement in task success and raised inference cost by over 20%. A July two-agent ablation reproduced the null correctness result for Claude Code and Codex CLI. The format is sticky because it is cheap to read, not because it has been shown to help. Sources: Virtualization Review (2026) (↗); arXiv (ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems) (2026) (↗); arXiv (2026) (↗)
The harness is a hidden variable in every benchmark. METR ran its GPT-5.6 Sol evaluation on its own ReAct harness and found the time horizon ranged from 11.3 hours to beyond 270 depending on how cheating was scored. A Terminal-Bench summary reports a model swap moved scores more than a scaffold swap (+52% against +17% relative), while automatic harness rewriting lifted a fixed model from 69.7% to 77%. The academic lane now names this the scaffold effect. Any ranking that does not state its harness is incomplete. Sources: METR (2026) (↗); jxzhangjhu blog (2026) (↗); Lil'Log (2026) (↗); arXiv (2026) (↗)
Multi-agent is in production without evidence it beats a single agent. Thoughtworks placed Coding agent swarms in Caution in April 2026 while Claude Code shipped Managed Agents subagent teams and obra/superpowers encodes dispatching-parallel-agents and worktrees as standard method. No retrieved source in any lane tests multi-agent against single-agent on cost or quality. The one nearby result, cross-model review of Claude by Codex and vice versa, concerns verification pairs rather than orchestration. Sources: Thoughtworks Technology Radar (2026) (↗); GitHub (2026) (↗); arXiv (2026) (↗)
Skills spread as a format in weeks and as a security problem in months. Registry listings went from 2,179 on 16 January to over 40,000 by 5 February 2026. Snyk's audit of 3,984 public skills found 36.82% with at least one flaw; a 98,380-skill analysis found 157 confirmed malicious; an earlier study put 26.1% of 31,132 skills as vulnerable. Skills are unsigned and rarely vetted, and Trail of Bits bypassed the scanners on ClawHub and skills.sh. Anthropic's donation of the standard to the Linux Foundation's Agentic AI Foundation on 22 October governs the format, not the supply chain. Sources: Firecrawl blog (2026) (↗); niteagent.com (2026) (↗); GitHub (2026) (↗); HiddenLayer (2026) (↗); Simon Willison's Weblog (2026) (↗); DEV Community (2026) (↗)
Portable in format, not in behaviour. Hysen Labs documents that obra/superpowers installs separately into 16 harnesses with no shared state. Each harness loads, triggers and sandboxes skills differently, so a "supports SKILL.md" claim means the folder is read, not that the agent behaves the same. The practitioner lane's judgement that skills resemble Cursor rules and MCP in being cheap to copy explains why value is settling in distribution and governance rather than in the files. Sources: Hysen Labs (2026) (↗); Thoughtworks Technology Radar (2026) (↗)
Effort is moving to review and verification, not generation. Böckeler's martinfowler.com piece frames the harness as feedforward guides and feedback sensors, and argues functional behaviour remains the unsolved part. Thoughtworks put Feedback sensors on Trial and Curated shared instructions on Adopt. Faros telemetry (vendor data) reports median PR review time up 441% against its 2025 baseline. DORA's 2026 report argues AI amplifies existing organisational quality, and a critic reads its own follow-up experiment as a 4% to 18% speedup, largely inconclusive. Sources: martinfowler.com (mirror) (2026) (↗); Thoughtworks Technology Radar (2026) (↗); Faros AI (2026) (↗); InfoQ (2026) (↗); makemeacto.cc (2026) (↗)
Labs are moving up the stack, and investors have noticed. Gartner names frontier labs competing directly with application vendors as the category's defining trend, with Anthropic, Cursor, GitHub and OpenAI as Leaders. Cursor's path from $29.3bn to a $60bn SpaceX acquisition in nine months shows the market pricing an independent harness as strategic rather than as a feature. Sources: Virtualization Review (2026) (↗); Bloomberg (2026) (↗); Wikipedia (2026) (↗
Evidence & Data
Measured, with a defined metric and a primary or near-primary source. Codex weekly users: over 2m (31 March), over 4m (21 April), over 5m (2 June 2026, about 20% knowledge workers). Later 10m, 20m and 25m figures use different or undefined windows and should not be chained on. Claude Code run rate: above $2.5bn in February 2026, from Anthropic's Series G announcement. Stack Overflow 2026: Claude Code 65.5% past-year use, up from 55% in the April pulse and 41% in 2025; agent use overall 59%, up from 31%; 73% of AI tool users use them daily. Gartner: $9.8bn to $11.0bn annualised as of April 2026. METR: public frontier 50% time horizon about 12 hours (range 5h to 61h), 80% horizon about 1.5 hours, suite unreliable above 16 hours. Sources: Gradually (2026) (↗); Constellation Research (2026) (↗); Constellation Research (2026) (↗); Stack Overflow (2026) (↗); Gartner (2026) (↗); METR (2026) (↗)
Measured but second-hand. Anthropic run rate above $30bn (April), $47bn (May) and $65bn (July 2026), all via aggregators citing Reuters and Bloomberg. Codex CLI at 99,624 GitHub stars (19 July 2026) and 41.8m npm downloads in May. Pi at 45,041 stars on 6 May 2026. Cursor ARR of roughly $4bn is an aggregator estimate, not a disclosure. Sources: Exploding Topics (2026) (↗); Panto (2026) (↗); Implicator.ai (2026) (↗); ValueAddVC (2026) (↗)
Narrative or untraced. A Claude Code figure of about $8bn for May 2026 appears only on a statistics blog. The "more than 70 tools support SKILL.md" tally comes from a single marketing post. A Databricks benchmark in which Pi had the highest pass rate could not be traced to Databricks. October comparison posts disagree on Terminal-Bench 4.0 (Codex 58.2% against Claude Code 57.9% within confidence intervals in one; 64.85% against 58.18% in another). Sources: AI Business Weekly (2026) (↗); lilachbullock.com (2026) (↗); explainx.ai (2026) (↗); Alongside (2026) (↗); Morph (2026) (↗)
Skills repositories. Star counts below are attention, not use. None is a primary GitHub reading except where a method is stated.
| Repository | Maintainer | Stars (date, source type) | What it encodes |
|---|---|---|---|
| obra/superpowers | Jesse Vincent | 304.5k (24 Oct 2026, star-history); 295,590 (5 Oct, GitHub API via practitioner post) | Brainstorm, write plan, execute plan, TDD, dispatch parallel subagents, git worktrees |
| mattpocock/skills | Matt Pocock | 302.7k (Oct 2026, parsed Trendshift listing; conflicts with other snapshots, unverified) | Not surfaced in sweep |
| anthropics/skills | Anthropic | 179,764 (5 Oct 2026, GitHub API via practitioner post); ~183k (cached GitHub page, undated) | Vendor reference skills and document-handling exemplars |
| ui-ux-pro-max-skill | Not surfaced | ~135k (12 Oct 2026, aggregator) | Front-end design and UX guidance |
| caveman | Not surfaced | ~111k (12 Oct 2026, aggregator) | Terse, token-sparing response style |
| addyosmani/agent-skills | Addy Osmani | 110.3k (Oct 2026, Trendshift); 101,480 (5 Oct, GitHub API via practitioner post) | Engineering practice and harness discipline |
| taste-skill | Not surfaced | ~95k (12 Oct 2026, aggregator) | Output quality and judgement heuristics |
| wshobson/agents | wshobson | 40,216 (5 Oct 2026, GitHub API via practitioner post) | Subagent role definitions |
Sources: Star History (2026) (↗); niteagent.com (2026) (↗); Trendshift (2026) (↗); jsdev.space (2026) (↗); GitHub (2026) (↗); GitHub (2026) (↗)
The pattern in the table is clear even if the numbers are soft. The most-starred skills encode process (plan first, test first, split work, verify) and register (be terse, have taste). Developers are not asking agents for more capability; they are asking them to behave like disciplined colleagues.
Signals & Tensions
Durable or transitional: the position. The bitter-lesson camp (Lee, Schmid, Bowne-Anderson, Arize) predicts task-specific scaffolding dissolves with each model release, and the ETH null result on context files supports them for the content of markdown directives. Osmani's counter, that harnesses move rather than shrink, is supported by Böckeler and by the fact that every lab is building more harness, not less. The reconciliation is that the layer has strata with different half-lives. Planning prompts, methodology reminders and generic skills are being absorbed into models and lab servers (Managed Agents, memory, compaction, tool search). Triggers, permissions, sandboxing, session persistence, scheduling and audit are economic and safety problems, not capability gaps, and no model absorbs them. Markdown wins as the interchange format for the durable stratum because it is the lowest-friction thing a human can review and a lab can parse. Invest in event plumbing, verification sensors and governance; do not invest in bespoke prompt methodology that the next model will ship for free. Sources: Han Lee blog (2026) (↗); Arize (2026) (↗); O'Reilly Radar (2026) (↗); arXiv (2026) (↗)
Minimalism against batteries-included is unresolved by data. Pi's stripped core and self-modifying design won an acquisition and a cult following, then added MCP within six months. OpenCode went the other way, into a background service with an SDK. Kingy.ai's fixed-model test (OpenCode 34/42, Claude Code 32/42, Codex 31/42) finds no significant gap, which suggests harness choice is about workflow fit rather than output quality. Sources: The Pragmatic Engineer (Substack) (2026) (↗); Kingy.ai (2026) (↗)
"Just markdown" is giving way to typed and tested skills, quietly. SkillsBench and the SoK paper on agentic skills are both preprints calling for replication, and a July study of SKILL.md "smells" treats the files as code with defects. The Snyk and Trail of Bits results make the case that unsigned prose with executable attachments is a supply chain, and supply chains get signing, scanning and evals. Expect schemas and verified bundles; expect the human-readable file to stay on top. Sources: arXiv (2026) (↗); arXiv (2026) (↗); arXiv (2026) (↗); arXiv (2026) (↗)
Overhyped: star counts. Underreported: the review bottleneck. Trendshift shows superpowers gaining 27,199 stars in 61 days, driven by social posts; clone counts include CI and mirrors. Meanwhile the only productivity signals in the sweep (DORA, Faros, METR's earlier developer study) point to verification as the constraint, and the harness features that address it (sensors, cross-model review, worktree isolation) get a fraction of the attention. Sources: Trendshift (2026) (↗); METR (2025) (↗)
Gartner's quadrant is contested at the detail level. Virtualization Review and a secondary write-up list different Leaders. Treat the market size as the reliable part and the placements as needing the primary document. Sources: Virtualization Review (2026) (↗); beri.net (2026) (↗)
Open Questions
Does multi-agent orchestration pay for itself? No retrieved source compares orchestrator-and-workers, verifier pairs or parallel worktrees against a single agent on cost or correctness, despite every major harness shipping the feature.
Why do context files fail controlled tests while practitioners swear by them? The ETH and July ablations measure task success on benchmarks; Böckeler's argument is about functional behaviour and team conventions that benchmarks do not score. Nobody has measured the latter. Sources: arXiv (ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems) (2026) (↗); martinfowler.com (mirror) (2026) (↗)
Did Anthropic reverse the third-party harness restriction, and on what terms? The brief asserts a reversal; the sweep found the April cut-off and nothing after it.
What are the true star counts? The mattpocock/skills 302.7k figure conflicts with every other snapshot, anthropics/skills ranges from 180k to 183k across undated pages, and no lane pulled the GitHub API directly for the full set.
Can time-horizon evidence still separate harness from model effects? METR says its suite cannot reliably measure above 16 hours and that cheating treatment swings results by an order of magnitude. Sources: METR (2026) (↗)
Where does margin settle? Labs own the subscription and can restrict rival harnesses at will; Cursor's $60bn exit says independent harnesses are worth owning; skills are free to copy. The VC lane found no a16z, Sequoia or Forrester thesis to adjudicate.
Who governs the skills supply chain now that the format belongs to a foundation? The AAIF owns the specification. Nobody owns signing, vetting or revocation, and the scanners have already been bypassed.
The year's lesson is that the harness did not get thinner as models improved. It got relocated: upward into lab servers, outward into triggers and schedulers, and sideways into markdown files that everyone reads and nobody has yet shown to help. Build on the plumbing, treat the prose as perishable, and keep a verifier in the loop.
Sources
Summary: ↑ Back to summary
Financial Press
25 sources
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| f1 | AI Coding Startup Cursor in Talks for About $50 Billion Valuation | Bloomberg | 2026-03 | Bloomberg report (primary financial press) that Cursor was seeking a round at about $50bn, nearly double its autumn 2025 mark, anchoring investor appetite for harness companies. |
| f2 | Anthropic's Claude Code revenue doubled since Jan. 1 | Constellation Research | 2026-02 | Relays Anthropic's February 2026 Series G disclosures: $14bn company run rate, Claude Code above $2.5bn (vendor-reported), 500+ customers over $1m a year. |
| f3 | Is Anthropic Profitable in 2026? Claude Code's Role in Q2 Growth | CodeConductor | 2026-08 | Aggregates a dated run-rate timeline ($14bn Feb, $30bn+ April, $47bn+ May, $65bn+ end July, attributed to Bloomberg and Reuters) and warns run rate is not recognised revenue; secondary, needs primary confirmation. |
| f4 | Anthropic Revenue and Valuation Statistics (October 2026) | Exploding Topics | 2026-10 | Secondary compilation citing Bloomberg, CNBC and Reuters for a $65bn run rate at end July 2026 and roughly $4.6bn of 2025 revenue per an IPO prospectus; a lead to chase, not a primary source. |
| f5 | Anthropic AI Statistics 2026: Users, Revenue & Market Share | Panto | 2026-05 | Records Claude Code at over $500m run rate by September 2025, then over $2.5bn by February 2026, and doubled weekly active users since 1 January 2026 (vendor-reported, relayed). |
| f6 | Claude Code Statistics 2026: Revenue, Market Share & Growth | AI Business Weekly | 2026-07 | Source of an unverified Claude Code figure of about $8bn ARR for May 2026 and a 54% market share claim; flagged here as untraceable to a primary source. |
| f7 | OpenAI Codex Statistics 2026: Data and Facts | Gradually | 2026-10 | Separates comparable Codex weekly-active-user figures (2m March, 4m April, 5m June 2026) from later 10m, 20m and 25m claims with undefined metrics. |
| f8 | OpenAI touts broadening Codex usage with 5 million weekly active users | Constellation Research | 2026-06 | Dated 2 June 2026 relay of OpenAI's vendor-reported 5m weekly Codex users, with knowledge workers about 20% of users. |
| f9 | Codex AI Statistics 2026: Users, Revenue & Growth | Panto | 2026-07 | Gives Codex CLI GitHub stars (99,624 on 19 July 2026) and npm downloads (41.8m in May 2026), plus 7m and 8m weekly users in July after the Codex and ChatGPT desktop merger; secondary, labelled as such. |
| f10 | Cursor in talks to raise $2B at $50B valuation after hitting $2B ARR in three years | The Next Web | 2026-03 | Records Cursor's $29.3bn Series D in November 2025 and the investor line-up (a16z, Thrive, Nvidia) for the next round, tracing the valuation path. |
| f11 | [Cursor (company)](https://en.wikipedia.org/wiki/Cursor_(company) | Wikipedia | 2026-08 | Summarises the April 2026 SpaceX option, the 16 June exercise and the 14 August 2026 close; a related Anysphere page names xAI as the April counterparty, so the acquirer entity should be checked against filings. |
| f12 | Cursor (Anysphere): Revenue, Funding & Valuation (2026) | ValueAddVC | 2026-08 | Estimates Cursor ARR at about $4bn (May 2026) and notes conversion to roughly 389 million SpaceX shares; an investor-analyst estimate, not a disclosure. |
| f13 | Anthropic cuts Claude subscribers off from OpenClaw in cost crackdown | The Next Web | 2026-04 | Documents the 4 April 2026 end of subscription coverage for third-party harnesses, the new pay-as-you-go tier, and Steinberger's reaction after joining OpenAI. |
| f14 | Anthropic blocks OpenClaw from Claude subscriptions | TechBuzz | 2026-04 | Attributes the April 2026 policy to emails reported by The Verge and describes the practical cost impact on OpenClaw-dependent workflows. |
| f15 | Tell HN: Anthropic no longer allowing Claude Code subscriptions to use OpenClaw | Hacker News | 2026-04 | Practitioner reaction thread (anecdotal) on subscription pricing at marginal cost and capacity constraints as the likely cause of the policy. |
| f16 | The Next AI Supply Chain Risk: Malicious Skills in Agentic AI | HiddenLayer | 2026-06 | Vendor security research stating that skills are not cryptographically signed and are rarely vetted, using OpenClaw as the case study and extending the risk to enterprise coding tools. |
| f17 | Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis | arXiv | 2026-04 | Academic paper stating that Agent Skills were adopted across Cursor, GitHub Copilot and Gemini CLI within months and lack mandatory marketplace review. |
| f18 | Agent Skills: Explore security threats and controls | Red Hat Developer | 2026-03 | Red Hat Developer guidance on malicious scripts, filesystem tampering and prompt injection in skills, an enterprise-vendor view of governance controls. |
| f19 | Agent Skills Threat Model | SafeDep | 2026-01 | Cites a study of 42,447 skills in which 26.1% of 31,132 analysed had at least one vulnerability, and lists typosquatting and hidden-injection threats. |
| f20 | awesome-agent-skills-security | GitHub | 2026-08 | Curated index citing a 98,380-skill registry analysis with 157 confirmed-malicious skills, and Trail of Bits scanner bypasses against ClawHub and skills.sh. |
| f21 | MalSkillBench: A Runtime-Verified Benchmark of Malicious Agent Skills | arXiv | 2026-06 | Shows skills persisting through a SessionStart hook written to Claude Code settings, linking skill security to the event-driven hooks layer of harnesses. |
| f22 | Claude AI Statistics 2026: Revenue, Users & Market Share | getpanto.ai | 102 days ago | Retrieved by this lane's web search. |
| f23 | Claude Code Is Doing $2.5B in Annualized Revenue - Just from the Terminal Tool | mindstudio.ai | 168 days ago | Retrieved by this lane's web search. |
| f24 | Claude Code Is Doing $2.5B in Annualized Revenue - Bigger Than Most Public SaaS Companies | mindstudio.ai | 169 days ago | Retrieved by this lane's web search. |
| f25 | Claude Code Usage Statistics 2026 (New Data) 📊 | serpsculpt.com | 156 days ago | Retrieved by this lane's web search. |
Frontier Lab & Model News
25 sources
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| t1 | Claude Code changelog - Claude Code Docs | Anthropic (Claude Code Docs) | 2026-10 | Official Anthropic changelog showing October 2026 additions (hooks, mods, unattended retry mode, sandbox and cloud-session changes) that trace Claude Code's move to event-driven, unattended workflows. |
| t2 | Claude Code Updates by Anthropic - October 2026 - Releasebot | Releasebot (changelog aggregator) | 2026-10 | Tracks releases 2.1.304 to 2.1.309 (Oct 2026), including Managed Agents subagent teams on multiagent_20261001, Claude Tag in Slack and Code Review controls, evidence of multi-agent and ticket-chat triggers moving into the harness. |
| t3 | Claude Code Changelog (October 2026) | gradually.ai | 2026-10 | Secondary changelog showing Claude Haiku 5.5 as default Haiku and an effort parameter on the Agent tool for sub-agents (release 2.1.292, 6 Oct 2026). |
| t4 | Claude Code Statistics 2026: Numbers and Facts | gradually.ai | 2026-10 | Aggregator compiling Claude Code run-rate milestones ($500M Aug 2025, $1bn Nov 2025) and Anthropic's 16 October 2026 claim of 5x commits per engineer; secondary, so figures need primary confirmation. |
| t5 | OpenAI Codex (Wikipedia) | Wikipedia | 2026-10 | Summarises reported growth from about 5% to roughly 40% of Claude Code's usage between Sept 2025 and Jan 2026 (anonymous sourcing) and Fortune's 1.6M WAU figure. |
| t6 | Google Moves Gemini CLI Into Antigravity CLI as Agent Platform Expands | Virtualization Review | 2026-05 | Reports Google's 19 May 2026 I/O announcement that Gemini CLI becomes Antigravity CLI, with skills, hooks and subagents retained, justified by a shift to multi-agent workflows. |
| t7 | Gemini CLI vs. Antigravity: What works, not the spec sheet | The New Stack | 2026-06 | Hands-on report that Gemini CLI (100,000+ GitHub stars) was replaced by a closed-source Go binary and stopped serving consumer accounts on 18 June 2026. |
| t8 | Google Antigravity (Wikipedia) | Wikipedia | 2026-05 | Describes Antigravity 2.0 (19 May 2026) as a standalone desktop app running parallel agents, custom subagent workflows and scheduled background tasks, alongside the CLI and SDK. |
| t9 | Antigravity CLI Prompting Guide: Google's Gemini CLI Replacement (2026) | SurePrompts | 2026-08 | States that GEMINI.md still works and AGENTS.md is read by Antigravity CLI, showing markdown directive files persisting across a harness rewrite. |
| t10 | SKILL.md in 2026: Authoring, Portability, Security | niteagent.com | 2026-10 | Gives dated GitHub API star counts (5 Oct 2026) for obra/superpowers (295,590), anthropics/skills (179,764) and addyosmani/agent-skills (101,480), and lists agentskills.io clients including Amp and OpenClaw; practitioner source, not primary. |
| t11 | What Are Agent Skills? The Open Standard Explained | lilachbullock.com | 2026-10 | Claims Anthropic donated Agent Skills to the Linux Foundation's Agentic AI Foundation on 22 October 2026 and that 70+ platforms support it; single marketing-blog source, needs primary confirmation. |
| t12 | Agent Skills as an Open Standard: How One Specification Conquered Every AI Coding Tool | Paperclipped | 2026-03 | Timeline of the 18 December 2025 spec release, rapid Microsoft and OpenAI adoption and 32 tools by March 2026, with a warning that security lagged adoption. |
| t13 | Agent Skills Explained: SKILL.md Format and Adoption (2026) | Atlan | 2026-09 | Reports 40+ clients on agentskills.io (checked September 2026), about 100 tokens of metadata per skill, and a 2026 study finding authoring flaws in 99%+ of SKILL.md files. |
| t14 | The Agent Skills Ecosystem in 2026: Who's Building, What's Working, and What's Next | Agentman | 2026-06 | Cites about 40 skills-compatible products in June 2026, SkillsMP indexing about 1.9 million public skills, and SkillsBench as a peer-reviewed skills benchmark; vendor blog with secondary citations. |
| t15 | What Are Agent Skills? The Complete 2026 Guide | Articsledge | 2026 | Relays a February 2026 Snyk audit of 3,984 public skills (36.82% with at least one flaw, 13.4% critical), the main security data point on community skills. |
| t16 | Frontier Risk Report (February to March 2026) | METR | 2026-05 | METR pilot with Anthropic, Google, Meta and OpenAI: public frontier 50% time horizon about 12h, internal models likely at 16h or more, and Time Horizon 1.1 near saturation. |
| t17 | Summary of METR's predeployment evaluation of GPT-5.6 Sol | METR | 2026-06 | Shows harness dependence: GPT-5.6 Sol's 50% time horizon is about 11.3h if cheating counts as failure but over 270h if counted as success, on METR's ReAct harness; links a Codex harness setup guide for third-party assessors. |
| t18 | METR | METR | 2026-10 | Index page listing METR's Risk Assessment of Claude Opus 5.5 (22 September 2026) and GPT-5.6 Sol (26 June 2026), the latest named-model evaluations. |
| t19 | METR Benchmarking | metavert.io | 2026-10 | Secondary summary of METR time-horizon caveats as of October 2026: Opus 4.6 corrected to about 12.0h and post-2023 doubling time of about 129 days. |
| t20 | Claude Mythos (Wikipedia) | Wikipedia | 2026-06 | Summarises Claude Opus 4.8 (28 May 2026) and Mythos 5 preview (9 June 2026) and US government access restrictions, model-side context for harness demand; Wikipedia, so verify against Anthropic. |
| t21 | Claude & Claude Code Updates: October 2026 Walkthrough | The AI Maker (Substack) | 2026-10 | Mentions Claude Duet, a proactive agent announced 21 October 2026 as a waitlist research preview, a sign of the shift from prompted to proactive agents. |
| t22 | Claude (language model) | en.wikipedia.org | Retrieved by this lane's web search. | |
| t23 | Articles | claude.com | 3 days ago | Retrieved by this lane's web search. |
| t24 | Anthropic Release Notes - October 2026 Latest Updates - Releasebot | releasebot.io | 4 days ago | Retrieved by this lane's web search. |
| t25 | ChatGPT | en.wikipedia.org | Retrieved by this lane's web search. |
Academic & arXiv
25 sources
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| a1 | Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? | arXiv (ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems) | 2026-02 | Gloaguen, Mündler, Müller, Raychev and Vechev (ETH Zurich, 2026) report no gain in task success from context files and a cost rise of over 20%, the central empirical challenge to the markdown-as-control-plane thesis. |
| a2 | Evaluating AGENTS.md (ETH SRI Lab publication page) | ETH Zurich SRI Lab | 2026 | Confirms the venue and the headline result (no improvement, over 20% extra inference cost) from the authors' own group page. |
| a3 | Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real Repositories | arXiv | 2026-07 | Independent July 2026 test on Claude Code (Sonnet 4.6) and Codex CLI (GPT-5.5) finding no detectable correctness effect from context-file strategies, with narrow process-level efficiency effects. |
| a4 | Probe-and-Refine Tuning of Repository Guidance for Coding Agents | arXiv | 2026-06 | Summarises the split in the literature: Lulla et al. report 28.6% less runtime with curated AGENTS.md while Gloaguen et al. report lower resolve rates, with agents following file instructions too literally. |
| a5 | Harness Engineering for Agentic AI Coding Tools: An Exploratory Study | arXiv | 2026-02 | First repository-level analysis (2,853 GitHub repos, eight configuration mechanisms, five tools) showing context files dominate and AGENTS.md acts as a cross-tool standard. |
| a6 | Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents, A Source-Code Study of Eleven Systems | arXiv | 2026-09 | Source-code comparison of eleven harnesses (including Claude Code, Codex, OpenCode, OpenHands) showing convergence on markdown context files and Codex adopting Claude Code's hook and plugin formats. |
| a7 | An Empirical Study of Harness Design for Coding Agents | arXiv | 2026-09 | Controlled harness ablation showing predefined tools help a weaker model while bash-only helps a stronger one (53% lower cost on SWE-Bench), evidence that scaffolding value depends on model strength. |
| a8 | The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable in Coding-Agent Evaluation | arXiv | 2026-07 | Holding the model fixed across Goose, OpenCode and OpenHands-SDK, harness choice moves tokens per solved task by up to 40x while pass-rate differences stay at 0 to 8 points (preliminary, ICML 2026 workshop submission). |
| a9 | Scanning the Harness: An Empirical Study of Supply-Chain Defects in AI Coding-Agent Configurations | arXiv | 2026-09 | Scans 3,171 public repositories of harness components (skills, commands, context files, plugin marketplaces) and notes AGENTS.md is stewarded by the Linux Foundation's Agentic AI Foundation. |
| a10 | SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks | arXiv | 2026-02 | First benchmark of Agent Skills (86 tasks, 11 domains, 7,308 trajectories) with no-skill, curated and self-generated conditions; curated skills lift pass rates but self-generated ones do not. |
| a11 | SkillsBench summary (alphaXiv overview) | alphaXiv | 2026-02 | Gives the per-condition numbers: fleet-wide pass rate from 33.9% to 50.5% with curated skills, and drops of 8.1 to 11.5 points with self-generated skills (the arXiv version's headline averages differ slightly, 16.2 versus 16.6 points). |
| a12 | SoK: Agentic Skills, Beyond Tool Use in LLM Agents | arXiv | 2026-02 | Systematisation paper that flags SkillsBench as a single, non-peer-reviewed source of quantitative skills evidence needing independent replication. |
| a13 | An Empirical Study of Agent Skills' Downstream Utility | arXiv | 2026-10 | October 2026 study over 87 SkillsBench tasks and a 37,596-skill corpus showing the same skills help some model-harness configurations and hurt others on 36.78% of tasks. |
| a14 | From Anatomy to Smells: An Empirical Study of SKILL.md in Agent Skills | arXiv | 2026-07 | Finds over 99% of SKILL.md files contain at least one authoring smell and that smells persist as skills evolve, quantifying the quality problem of unstructured markdown skills. |
| a15 | An Empirical Study of Downstream Adaptation for Agent Skills | arXiv | 2026-07 | First study of how public skills are forked and adapted, and the source describing the SKILL.md structure (YAML front matter, Markdown body, bundled scripts and resources). |
| a16 | Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale | arXiv | 2026-01 | Early-2026 large-scale audit of 31,132 skills from two marketplaces finding 26.1% with at least one vulnerability across 14 patterns. |
| a17 | Malicious Agent Skills in the Wild: A Large-Scale Security Empirical Study | arXiv | 2026-02 | Behaviourally verified 98,380 community skills, confirming 157 malicious ones (632 vulnerabilities), with one actor behind 54.1% of cases and 93.6% removal within 30 days of disclosure. |
| a18 | Technical Report: Exploring the Emerging Threats of the Agent Skill Ecosystem | arXiv | 2026-05 | Practitioner-authored report finding 13.4% of skills with critical issues and up to 9% of popular skills.sh skills exhibiting attack-vector patterns, though none confirmed malicious. |
| a19 | SkillSieve: A Hierarchical Triage Framework for Detecting Malicious AI Agent Skills | arXiv | 2026-04 | Evaluates a detector on 49,592 ClawHub skills and records that OpenClaw's marketplace audits found 13% to 26% of skills vulnerable, linking the OpenClaw skills market to the security literature. |
| a20 | Towards a Risk Assessment of Malicious Skill Files in Coding Agents | arXiv | 2026-08 | Releases 2,826 adversarial skill files mapped to MITRE ATT&CK; also repeats an unverified claim of 162,000 GitHub stars for a popular skills repository in three months, a figure to trace before reuse. |
| a21 | Metrics of Agent Ability | METR | 2026-07 | METR note (July 2026) on moving from binary time horizons to continuous, human-grounded metrics as the existing suite saturates. |
| a22 | Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity | METR | 2025-07 | METR's RCT (16 developers, 246 tasks, mostly Cursor Pro with Claude 3.5/3.7) found a 19% slowdown; METR now marks it out of date and points to an early-2026 continuation, so it is a stale baseline. |
| a23 | The Rise of AI Teammates in Software Engineering (SE 3.0): How Autonomous Coding Agents Are Reshaping Software Engineering | arXiv | 2025-07 | Introduces AIDev, the baseline measured dataset of agent-authored GitHub PRs (Codex, Devin, Copilot, Cursor, Claude Code); a snapshot to August 2025 and so predates the 2026 harness changes. |
| a24 | Security in the Age of AI Teammates: An Empirical Study of Agentic Pull Requests on GitHub | arXiv | 2026-01 | Uses the AIDev curated subset of 33,596 PRs and shows the agent mix (Codex 21,799; Claude Code 459), a measured caution that PR-count data skews towards bot-labelled agents. |
| a25 | Cross-Model LLM Code Review: Should you use Claude to review Codex or vice versa? | arXiv | 2026-07 | Tests writer-reviewer pairs across Claude Code (Opus 4.7) and Codex CLI (GPT-5.5) with cost accounting, direct evidence on verifier-pair multi-agent patterns. |
VC & Analyst Reports
25 sources
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| v1 | Leading in Enterprise AI Coding Agents Requires More Than Product Momentum (Enterprise AI Coding Agents: 2026 Market Guide and Trends) | Gartner | 2026-05 | Primary Gartner page sizing the enterprise AI coding agent market at about $9.8bn to $11.0bn annualised (April 2026) and naming labs moving up the stack and vertically integrated versus model-agnostic offerings as defining shifts. |
| v2 | Leading in Enterprise AI Coding Agents Requires More Than Product Momentum (Gartner Insights abstract) | Gartner | 2026-05 | Gartner's own record of the 20 May 2026 publication date and its framing of the Magic Quadrant for Enterprise AI Coding Agents through an enterprise buying lens. |
| v3 | AI Firms Push Cloud Giants from 'Leaders' Quadrant in Gartner AI Coding Report | Virtualization Review | 2026-06 | Trade coverage of the first Gartner Magic Quadrant for this category: Anthropic, Cursor, GitHub and OpenAI as Leaders, hyperscalers as Challengers, and cost governance as the main ROI question (secondary reporting of a Gartner document). |
| v4 | Gartner: AI Coding Agents Hit $10B, IDEs Going Optional | Enterprise DNA | 2026-05 | Secondary summary carrying Gartner's prediction that by 2027 over 65% of agentic-coding teams will treat the IDE as optional, a forecast, not a measurement. |
| v5 | Gartner's 2026 AI Coding Agent Report: What It Actually Means for Your Build Decision | Zetrix | 2026-08 | Secondary source attributing to Gartner a 19.3% net average productivity gain and 90% of engineering leaders reporting improved output; the figures are not traced to the primary report here. |
| v6 | Gartner 2026 MQ: Enterprise AI Coding Agents Leaders and New Criteria | Creeta News | 2026-05 | Secondary account that Cursor scored furthest on Completeness of Vision (in-house Composer model, SDK-driven scheduled workflows) while GitHub Copilot scored highest on Ability to Execute. |
| v7 | Gartner's 5 AI Coding Agent Leaders: A 17x ROI Guide | beri.net | 2026-05 | Conflicting secondary account listing different Leaders (GitHub, Amazon, Cognition, GitLab, Google Cloud) and a $3.0bn to $3.5bn early-2025 market baseline; useful as a warning that Gartner quadrant details are repeated inconsistently. |
| v8 | Claude Statistics 2026: Revenue, Users, and Market Share | Growthcentr | 2026-10 | Secondary compilation putting Anthropic's run rate at $65bn by end of July 2026 (said to be investor-shared figures confirmed by CNBC) and noting $2.5bn remains the latest official Claude Code figure; the $65bn figure is not primary here. |
| v9 | Claude AI Stats October 2026 | fatjoe | 2026-10 | Anecdotal secondary source claiming 2 million or more Claude Code weekly active users (May 2026) and third-party estimates near $8bn run rate; flagged as untraceable to an Anthropic primary source. |
| v10 | Claude Code Hits $2.5 Billion Run Rate in 2026: What It Tells B2B Buyers | LinkedOtter | 2026-06 | Secondary source citing OpenAI's reported 5 million Codex weekly active users as of 2 June 2026, the only Codex usage figure found, and contrasting user counts with revenue depth. |
| v11 | Best AI Coding Agents in 2026: Harness, Cost, and Accuracy Compared | Firecrawl | 2026-06 | Reports OpenCode at over 171,000 GitHub stars and 1.68 million weekly npm downloads (June 2026) and Cursor's $2.3bn Series D at $29.3bn on over $1bn annualised revenue (November 2025); installs are not users. |
| v12 | [Cursor (code editor)](https://en.wikipedia.org/wiki/Cursor_(code_editor) | Wikipedia | 2026-09 | Records the April 2026 New York Times report of a SpaceX option to acquire Cursor for $60bn and the November 2025 Series D co-led by Thrive and Coatue with a16z participating. |
| v13 | OpenCode vs OpenClaw (2026): Which Should You Run? | Agent37 | 2026-09 | Dated star counts (19 September 2026): OpenCode above 208,000 and OpenClaw above 390,000, with OpenClaw under a US non-profit foundation; stars measure attention, not active users. |
| v14 | Agent Harness 2026: OpenClaw vs Hermes vs Pi Compared | claw4science | 2026-05 | Star counts verified against GitHub on 29 May 2026 (OpenClaw 375,222; Pi 57,298) and the point that OpenClaw embeds Pi's SDK, so the two are not like-for-like. |
| v15 | OpenClaw | Wikipedia | 2026-09 | Notes the 30 August 2026 OpenClaw 2.0 release by the OpenClaw Foundation, with The Register flagging that shared sessions lack network and file-system security boundaries. |
| v16 | OpenClaw Statistics 2026: Users, Growth, Revenue & Adoption | Panto | 2026-09 | Growth timeline for OpenClaw (launched November 2025 as Clawd, 100,000 stars by late January 2026, about 369,000 by May 2026); aggregator, treat as lead. |
| v17 | Agent Skills: Anthropic's Next Bid to Define AI Standards | The New Stack | 2025-12 | Contemporaneous report of the December 2025 open-standard release, framed by The New Stack as a repeat of the MCP playbook. |
| v18 | Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward | arXiv | 2026-02 | Academic survey recording over 62,000 stars on anthropics/skills within four months of the open standard, and covering skill security and acquisition methods. |
| v19 | GitHub: anthropics/skills | GitHub | 2026 | Primary repository for the official skills set; the cached view shows about 183k stars, but the snapshot date is unclear and conflicts with lower early-2026 counts, so the figure needs a dated re-check. |
| v20 | Best AI Coding Agent Harness in 2026: Claude Code vs Codex vs OpenCode | Alongside | 2026-10 | October 2026 comparison reporting a statistical tie on Terminal-Bench 4.0 (Codex 58.2% versus Claude Code 57.9%) and noting that the benchmark scores the harness and model as a pair. |
| v21 | The 2026 AI Coding Agent Landscape: Leaders, Costs, Harness | Capital and Compute | 2026-08 | Terminal-Bench 2.1 (June 2026) scores by agent and model pair and the argument that the harness matters as much as the model. |
| v22 | What's the Best Coding Agent? 2026 Edition | DoltHub | 2026-08 | Practitioner account (DoltHub, August 2026) of models and harnesses being trained together, then loosening after the November 2025 Opus release drove mass adoption and pushed OpenAI to improve Codex. |
| v23 | Best AI Coding Agents in 2026: I Tested Claude Code, Codex, OpenCode, and Pi for 14 Days | Tensorlake | 2026-08 | Small vendor benchmark (30 tasks) in which Pi, with four tools, matched Claude Code and Codex on pass rate at lower cost; a single-author test and weak evidence. |
| v24 | Claude Code vs Codex: Which AI Coding Agent Should You Use in 2026? | firecrawl.dev | 144 days ago | Retrieved by this lane's web search. |
| v25 | Best AI Coding Agents in 2026: Claude Code vs Codex vs Cursor | turingcollege.com | 58 days ago | Retrieved by this lane's web search. |
Blogs & Independent Thinkers
25 sources
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| b1 | Donating Agent Skills to the Agentic AI Foundation | Simon Willison's Weblog | 2026-10 | Dated primary-adjacent record (22 October 2026) that Anthropic is donating the SKILL.md standard to the AAIF under the Linux Foundation, plus Willison's December 2025 prediction of exactly this. |
| b2 | Simon Willison on skills (tag page) | Simon Willison's Weblog | 2026-01 | Running log of 15 posts on skills, including the December 2025 note that OpenAI added Skills to Codex docs and the Agent Skills homepage. |
| b3 | Anthropic Just Gave Away Agent Skills | DEV Community | 2026-10 | Practitioner reading of the AAIF donation: governs the format, not the supply chain, so vetting community skills stays the user's problem. |
| b4 | Agent Harness Engineering (Addy Osmani blog) | addyosmani.com | 2026-04 | Defines harness engineering, credits Trivedy with the term, and maps each harness feature (compaction, skills, bash) to a model weakness. |
| b5 | Agent Harness Engineering (O'Reilly Radar repost) | O'Reilly Radar | 2026-05 | Osmani's claim that harnesses move rather than shrink as models improve, directly relevant to whether scaffolding stays necessary. |
| b6 | Hidden Technical Debt of AI Systems: Agent Harness | Han Lee blog | 2026-05 | Applies the Bitter Lesson to harness code, arguing structure fitted to 2026 capability is what the next model release dissolves. |
| b7 | Self-Evolving Agentic Harnesses | jxzhangjhu blog | 2026 | Cites Terminal-Bench figures showing model swaps beat scaffold swaps (+52% against +17% relative) alongside automatic harness rewriting lifting 69.7% to 77%. |
| b8 | Harness Engineering for Self-Improvement | Lil'Log | 2026-07 | Lilian Weng's July 2026 survey of propose-evaluate-accept loops where agents improve their own harness. |
| b9 | Language model harnesses are compositional generalizers | alexzhang13.github.io | 2026-07 | July 2026 essay by Alex Zhang and Omar Khattab warning that hand-built harness structure risks running afoul of the bitter lesson within months. |
| b10 | Are Agent Harnesses Dying? What Distillation Changes | Arize | 2026-09 | Recent analysis of models absorbing harness behaviour via training on corrected runs, with a cited 21 point task-success gain and a pointer to TheSequence's 'Great Absorption' thesis. |
| b11 | AI Agent Harness, 3 Principles for Context Engineering, and the Bitter Lesson Revisited | Hugo Bowne-Anderson (Substack) | 2025-12 | Substack piece by Hugo Bowne-Anderson arguing harness assumptions go obsolete in about six months and should be stripped back as models improve. |
| b12 | The importance of Agent Harness in 2026 | philschmid.de | 2026-01 | Phil Schmid's examples of rebuilds (Manus five refactors in six months, Vercel removing 80% of tools) as anecdotal evidence for lightweight harnesses. |
| b13 | Why Most Agent Harnesses Are Not Bitter Lesson Pilled | X (article post) | 2026-01 | Argues workflows and fixed subagent graphs plateau, and names a countermovement favouring a loop with good tool design; a single social post, so lower evidential weight. |
| b14 | Pi Coding Agent Turns Minimalism Into Harness Rebellion | Implicator.ai | 2026-05 | Gives dated Pi figures (45,041 stars and 2,143 packages on 6 May 2026), the Earendil acquisition of 8 April 2026, and comparison star counts for Codex and Gemini CLI. |
| b15 | Building Pi, and what makes self-modifying software so fascinating | The Pragmatic Engineer (Substack) | 2026-03 | Pragmatic Engineer newsletter interview recording that Pi was built because Claude Code became unpredictable and that OpenClaw is built on Pi. |
| b16 | Pi Agent Harness (pi.dev): Minimal Coding Agent by Mario Zechner (2026) | explainx.ai | 2026-10 | Reports Pi 1.0 on 2 October 2026, MCP entering Pi core on 29 September, and an unverified Databricks benchmark claim that Pi led on pass rate. |
| b17 | obra/superpowers: 16 install paths, one methodology | Hysen Labs | 2026-10 | Records 296,826 stars (synced 9 October 2026) and the point that Superpowers installs separately into 16 harnesses, limiting portability in practice. |
| b18 | obra/superpowers GitHub trending stats (Trendshift) | Trendshift | 2026-10 | Shows +27,199 stars in 61 days to 25 October 2026, and quotes the author's unaudited clone claims (56M+ clones in two weeks). |
| b19 | obra/superpowers Skills Collection | claudeskills.info | 2026-07 | Lists 14 skills, 287,886 stars and 3.2M total installs (last update 24 July 2026), with per-skill install counts; directory telemetry, not active users. |
| b20 | using-superpowers (skills.sh) | skills.sh | 2026-10 | Directory page showing 271.0K installs and 304.6K stars, first seen 19 January 2026, illustrating how skills.sh reports installs. |
| b21 | Agent Skills Explained: How SKILL.md Files Work and Why They're Everywhere | Firecrawl blog | 2026-03 | Cites the Bosch and CMU arXiv study (2602.08004) of 40,285 skills and the 18.5x growth in 20 days, and lists the harnesses that adopted SKILL.md. |
| b22 | From Skills to Agents: Bridging Claude Skills and AGENTS.md | dave.engineer | 2025-11 | Baseline from November 2025 when SKILL.md worked only in Claude; shows the early workaround of compiling skills into AGENTS.md. Stale for capability claims. |
| b23 | Claude Code, Codex CLI, OpenCode Update in 4 Days | Shattered | 2026-10 | Dates the 12 to 16 October 2026 release cluster (OpenCode 2.0 as a background service with API and SDK, Codex 0.163.0 mid-task steering, Claude Code forked subagents). |
| b24 | OpenCode Review 2026: Benchmarks, Safety & Rivals | Kingy.ai | 2026-10 | Fixed-model test (42 tasks) finding no statistically significant solve-rate difference between OpenCode, Claude Code and Codex CLI. |
| b25 | Claude Code Alternatives (2026): 11 Tested | Morph | 2026-10 | Gives OpenCode 214K against Claude Code 151K GitHub stars (October 2026), a star metric and not active users. |
Tech Industry & Practitioner
25 sources
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| p1 | April 2026 Technology Radar (Vol. 34) | Thoughtworks Technology Radar | 2026-04 | Primary Thoughtworks document that names harness engineering as a theme and rates Agent Skills (Trial), Curated shared instructions (Adopt), Coding agent swarms and OpenClaw (Caution). |
| p2 | Agent Skills (Technology Radar blip) | Thoughtworks Technology Radar | 2026-04 | Official radar entry, published 15 April 2026, defining Agent Skills as an open standard for just-in-time context and linking it to MCP by default and instruction bloat cautions. |
| p3 | Technology Radar: Tools | Thoughtworks Technology Radar | 2026-04 | States that Claude Code introduces features other tools follow (skills, subagents, remote control) and that Cursor appears alongside Claude Code as a default for Thoughtworks delivery teams. |
| p4 | Technology Radar: Techniques | Thoughtworks Technology Radar | 2026-04 | Shows the Adopt, Trial and Caution placement of curated instructions, feedback sensors and swarms, plus the advice to track collaboration quality at team level alongside DORA metrics. |
| p5 | Technology Radar Vol. 33 (Nov 2025) | Thoughtworks Technology Radar | 2025-11 | Baseline edition from the start of the window, covering AGENTS.md, anchoring coding agents to a reference application and the Team of coding agents technique. |
| p6 | As AI Accelerates Software Complexity, Thoughtworks Technology Radar Urges a Return to Engineering Fundamentals to Combat Cognitive Debt | Thoughtworks | 2026-04 | Thoughtworks CTO Rachel Laycock's framing that humans must implement practices and technical harnesses to use rising AI capability safely. |
| p7 | Harness engineering for coding agent users (summary of Birgitta Böckeler, martinfowler.com, 2 April 2026) | martinfowler.com (mirror) | 2026-04 | Summarises the guides-and-sensors model and the claim that functional behaviour remains unsolved and needs human judgement; this is a third-party mirror of the martinfowler.com article, so check the original. |
| p8 | Martin Fowler on X: Böckeler harness engineering post announcement | X (Martin Fowler) | 2026-04 | Dated announcement (2 April 2026) confirming the article's authorship and its follow-up to an earlier memo. |
| p9 | Harness Engineering for Coding Agent Users: Guide | beri.net | 2026-08 | Gives the publication chronology (memo 17 February 2026, full article 2 April 2026) and attributes the article to Böckeler as a Thoughtworks Distinguished Engineer; secondary source. |
| p10 | Harness Engineering for Coding Agents: What It Is (2026) | Agentastic | 2026-10 | Traces the term to OpenAI's 11 February 2026 Codex post and describes worktree-per-task parallel agent practice; vendor-adjacent explainer, useful as a lead. |
| p11 | AI coding agents: Stack Overflow Developer Survey 2026 data | Stack Overflow | 2026-10 | Primary survey table giving usage shares for Claude Code (65.5%), Copilot, Codex, Cursor, Antigravity, OpenCode, Pi and Aider (survey measured, 'used in the past year'). |
| p12 | AI data 2026 (Stack Overflow Developer Survey) | Stack Overflow | 2026-10 | Primary data on AI tool usage frequency, with 73% of coding assistant and agent users reporting daily use and 26.2% using agents or automated workflows. |
| p13 | The results of the 2026 Developer Survey are here! | Stack Overflow Blog | 2026-10 | Stack Overflow's own headline summary, including the context and trust findings (source attribution 93%) relevant to markdown context files and skills. |
| p14 | Getting ready for 2026 results: A look back on Developer Survey findings | Stack Overflow Blog | 2026-09 | Gives the April 2026 pulse survey trend: agent use 31% in 2025 to 59%, Claude Code 41% to 55%, with enthusiasm cooling. |
| p15 | Stack Overflow 2026 Survey: VS Code Leads, Visual Studio Takes Third | Visual Studio Magazine | 2026-10 | Independent reporting of the 12,255-respondent agent question and the editor result in which Cursor was used regularly by 14.9%, useful for separating tool-used from regular-use metrics. |
| p16 | New DORA Report Claims Strong Engineering Foundations Drive AI Return on Investment | InfoQ | 2026-05 | InfoQ coverage of DORA's ROI of AI-Assisted Software Development (2026.01), with Nathen Harvey's argument that organisational systems, not tools, drive returns. |
| p17 | What should one think of the 2026 DORA brochure? | makemeacto.cc | 2026-06 | Named critique of the DORA ROI report's evidence quality, including its follow-up experiment showing a 4% to 18% speedup that it calls largely inconclusive; blog commentary, not peer review. |
| p18 | Key Takeaways from the DORA Report 2025: How AI is Reshaping Software Development Metrics | Faros AI | 2026 | Faros vendor telemetry comparing 2025 and 2026 findings (for example review time +441%, incidents per PR +242.7%), flagging a review bottleneck; vendor-reported. |
| p19 | obra/superpowers (GitHub) | GitHub | 2026 | Primary repository showing its skills (brainstorming, writing-plans, executing-plans, parallel agents, worktrees, code review) and about 305k stars, encoding a plan-then-execute, subagent-driven method. |
| p20 | obra/superpowers: 304.5k Stars, Global Rank #13 | Star History | 2026-10 | Dated star snapshot (24 October 2026) and global GitHub rank list; a third-party measurement of repository stars. |
| p21 | The best Claude skills in 2026: top picks by use case | Walma AI | 2026-10 | Aggregator snapshot of stars on 12 October 2026 for superpowers (about 298k), anthropics/skills (about 181k), ui-ux-pro-max-skill (about 135k), caveman (about 111k) and taste-skill (about 95k); a lead, not primary evidence. |
| p22 | 30 Popular AI Agent Skills Ranked by GitHub Stars | jsdev.space | 2026-10 | Candidly notes that stars attach to repositories rather than skills and that most counts are provisional snapshots, which supports caution on repeated star tallies. |
| p23 | Thoughtworks Technology Radar Vol. 34, Agentic AI Digest | Agentic AI Knowledge Base | 2026-04 | Third-party digest that restates the radar's supply chain caution on unreviewed third-party skills and the Agent Skills Trial entry; useful for cross-checking, but check against the primary radar. |
| p24 | Technology Radar | thoughtworks.com | 354 days ago | Retrieved by this lane's web search. |
| p25 | Thoughtworks Tech Radar Recap: The AI Refresher | tanatloke.medium.com | 180 days ago | Retrieved by this lane's web search. |