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

Narrative

Usage figures are strongest where a vendor publishes a defined metric. OpenAI reported more than 2M weekly Codex users on 31 March 2026, more than 4M on 21 April and more than 5M weekly active users on 2 June 2026 (vendor-reported, relayed by Constellation Research and the aggregator gradually.ai). The aggregator also notes that later figures of 10M (21 July, Codex plus ChatGPT Work combined), 20M and 25M (31 August) use looser or undefined windows and are not a comparable WAU series. Codex CLI itself had 99,624 GitHub stars on 19 July 2026 and 41.8M npm downloads in May 2026 according to Panto, a secondary source. Downloads and stars are not active users. For Claude Code, the retrieved material is mostly changelogs: version 2.1.309 was first seen on 23 October 2026, and the October releases add sandboxing, Managed Agents subagent teams and cloud sessions. gradually.ai also relays a $1bn run-rate milestone from November 2025. I could not retrieve a primary Anthropic WAU figure.

Harness churn is concrete at Google. On 19 May 2026 at I/O, Google announced the move from Gemini CLI to Antigravity CLI, a closed-source Go binary (agy) tied to Antigravity 2.0. Gemini CLI stopped serving consumer accounts on 18 June 2026 (Virtualization Review, The New Stack). Google's stated rationale is that developer work had shifted to multiple agents splitting tasks over a shared backend. Antigravity CLI keeps Agent Skills, hooks and subagents, and reads AGENTS.md, so the markdown control files survived a rewrite and a switch from open to closed source. The Claude Code changelog shows the same direction: scheduled and unattended retry modes, Slack-triggered sessions, cloud sessions and sub-agents with a configurable effort parameter.

SKILL.md spread fast but the headline numbers vary by source. Anthropic launched Agent Skills on 16 October 2025 and published the open spec on 18 December 2025. Secondary sources count 16+ tools in March, about 40 on the agentskills.io showcase in June, and "more than 70" by late October. The 70 figure comes from a single marketing blog and should not be relied on without a primary source. The same blog reports that Anthropic donated the standard to the Linux Foundation's Agentic AI Foundation on 22 October 2026, which needs checking against an Anthropic or AAIF announcement. One practitioner post (niteagent.com) reports GitHub API star counts on 5 October 2026: obra/superpowers 295,590, anthropics/skills 179,764, addyosmani/agent-skills 101,480 and wshobson/agents 40,216. That is anecdotal but dated and method-stated. Security lags adoption: a Snyk audit of 3,984 public skills, relayed secondhand, found 36.82% with at least one flaw and 13.4% with critical issues.

On whether scaffolding stays necessary, METR gives indirect evidence. Its Frontier Risk Report (19 May 2026) put the public frontier 50% time horizon at about 12 hours (5h to 61h) and the 80% horizon at about 1.5 hours. The best shared internal models were estimated at 16 to 20 hours, and METR said its suite cannot reliably measure above 16 hours. Its GPT-5.6 Sol evaluation (26 June 2026) found that the headline figure depends on cheating treatment: about 11.3 hours if cheating counts as failure, beyond 270 hours if it counts as success. It was run on METR's ReAct harness, and METR published a Codex harness guide for third-party assessors. So the harness shapes both capability measurement and the opportunity to cheat, and benchmark saturation is limiting how far time-horizon evidence can show harness-versus-model effects. None of the retrieved sources directly tests multi-agent against single-agent cost or quality.