The Agent Harness Landscape and the Shift to Markdown-Defined, Event-Driven Agents

Narrative

The strongest practitioner signal in this lane is that "harness" has become the unit of analysis. Thoughtworks Technology Radar Vol. 34 (April 2026) made "putting coding agents on a leash" a headline theme, and placed Agent Skills (Trial, new), Feedback sensors for coding agents (Trial) and Curated shared instructions (Adopt) on the radar. It put Coding agent swarms and OpenClaw in Caution, and the radar's tools page says Claude Code keeps introducing features, such as skills, subagents and remote control, that other tools then follow. Birgitta Böckeler's martinfowler.com article (memo 17 February, full piece 2 April 2026) supplies the vocabulary: feedforward guides and feedback sensors, computational versus inferential controls. She argues that functional behaviour is the unsolved part and still needs human judgement. That cuts against the view that scaffolding disappears as models improve. Vol. 33 (November 2025) is the baseline, when the radar was still discussing AGENTS.md and a "team of coding agents" as a technique.

Hard usage data is thin and mostly survey-based. The Stack Overflow 2026 Developer Survey, published 6 October 2026, has the best like-for-like numbers. Among 12,255 respondents to the coding agent question, 65.5% had used Claude Code and 58.7% Copilot. OpenAI Codex scored 29.5%, Cursor 25.5%, Google Antigravity 16.0%, OpenCode 13.6% and Pi 4.4%. Aider scored 2.1%. These are "used in the past year" shares, not active users, and the list includes IDE assistants as well as terminal harnesses. Stack Overflow's own pulse survey in April 2026 had Claude Code at 55%, up from 41% in the 2025 survey, and agent use overall at 59%, up from 31%. The survey shows 73% of AI coding tool users use them daily. I found no primary weekly-active-user or API-traffic figures for any harness, so none are reported here.

Skills repositories show how developers want agents to behave, but star counts need care. obra/superpowers showed 304.5k stars on star-history on 24 October 2026, and its GitHub page lists skills such as brainstorming, writing-plans, executing-plans, dispatching-parallel-agents and git worktrees, so it encodes a plan-then-execute, subagent-driven method with TDD. anthropics/skills showed about 183k stars on a cached GitHub page of unclear date. A 12 October 2026 aggregator snapshot gives ui-ux-pro-max-skill about 135k, caveman about 111k and taste-skill about 95k. Trendshift lists addyosmani/agent-skills at 110.3k and a mattpocock/skills entry at 302.7k. That last figure is parsed from a listing and conflicts with other snapshots, so verify it against the GitHub API before use. Stars measure attention, not installs or use. Thoughtworks warns that unreviewed third-party skills carry supply chain risk.

Productivity evidence is contested and often vendor-driven. DORA's "ROI of AI-assisted Software Development" (2026.01, covered by InfoQ in May 2026) argues that AI amplifies existing organisational quality, and that internal platform quality and clear workflows drive returns. A critical blog post calls the report a brochure, and says the follow-up experiment it describes showed a speedup of 4% to 18% that is largely inconclusive. Faros telemetry, which is vendor data, reports deeper review bottlenecks in 2026, with median PR review time up 441% against its 2025 baseline. Taken together, the sources suggest that the review and verification layer, not code generation, is where harness effort is going.

This lane did not retrieve primary material on the Anthropic third-party harness ban and its reversal, Codex plan bundling, event-driven routines, or the Antigravity CLI transition. Those need vendor changelog sources from other lanes.