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

Narrative

Independent commentary in this lane treats the harness as the main object of engineering. Addy Osmani (blog, 19 April 2026, reposted by O'Reilly Radar on 15 May 2026) credits Viv Trivedy with coining "harness engineering" and argues that a decent model with a great harness beats a great model with a bad harness. His O'Reilly piece adds that harnesses "move" rather than shrink as models improve. Simon Willison's definition (an agent runs tools in a loop to achieve a goal) is the working consensus. On the open question of absorption, Lee Han Chung (8 May 2026), Phil Schmid, Hugo Bowne-Anderson (Substack) and Arize (about late September 2026) all apply Sutton's Bitter Lesson and predict that task-specific scaffolding gets dissolved by the next model release. A Terminal-Bench summary (jxzhangjhu blog, 2026) points the same way, reporting that a model swap moved scores more than a scaffold swap (+52% against +17% relative). Counter-evidence also appears: Lilian Weng (July 2026) and the same blog describe automatic harness rewriting lifting a fixed model from 69.7% to 77% on Terminal-Bench. The evidence therefore favours "specific tricks decay, the harness layer persists", not either extreme.

On skills, the main dated facts come from Simon Willison's weblog. On 22 October 2026 he recorded Anthropic's donation of the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, and noted he had predicted this in December 2025. He also logged OpenAI's Codex adopting skills in December 2025. A DEV Community post (23 October 2026) reads the donation as format governance only, not supply-chain governance, and stresses that skills are markdown files agents are built to trust. Firecrawl's explainer cites an arXiv study (2602.08004, Bosch Research and Carnegie Mellon, February 2026) counting 40,285 listed skills, a rise from 2,179 on 16 January to over 40,000 by 5 February. That is a registry listing count, not usage. Hysenlabs observes that obra/superpowers installs separately into 16 harnesses with no shared state, which is a practical limit on portability.

Star counts for obra/superpowers vary by date and source, and none is a primary GitHub reading. Retrieved figures are 89K+ (Termdock, March 2026), 287,886 (claudeskills.info, last updated July 2026), 296,826 (Hysenlabs, synced 9 October 2026) and 304.6K with 271.0K installs of one skill (skills.sh page). Trendshift shows +27,199 stars in 61 days to 25 October 2026, and it quotes the author's own post about 300K stars and clone counts, which is vendor-reported and unaudited. Skills.sh install counts are directory telemetry, not active users. On harnesses, a Pi write-up by Implicator (May 2026) reports 45,041 stars for Pi on 6 May, the 8 April acquisition of Pi by Armin Ronacher's Earendil, and a stripped core with no MCP, subagents or plan mode. Explainx (October 2026) reports Pi 1.0 on 2 October and MCP entering Pi's core on 29 September. It also cites a Databricks benchmark in which Pi had the highest pass rate, which I could not trace to a primary Databricks document. Morphllm's October 2026 alternatives list gives OpenCode 214K stars against 151K for Claude Code. Firecrawl quotes OpenAI's 1 June 2026 announcement of more than 5 million weekly Codex users, a vendor-reported figure.

Comparison posts from October 2026 (Morph, Alongside, Kingy.ai, Shattered) agree that Claude Code, Codex and OpenCode shipped updates between 12 and 16 October 2026, with OpenCode 2.0 repositioned as a background service with an API and SDK. They split on rankings. Alongside reports Terminal-Bench 4.0 pairs within confidence intervals of each other (Codex 58.2%, Claude Code 57.9%), while Morph reports a 64.85% against 58.18% gap. Kingy.ai's fixed-model test (OpenCode 34/42, Claude Code 32/42, Codex 31/42) shows no significant differences. These are SEO-adjacent sources whose benchmark claims should be treated as leads, and the model names they cite could not be checked here. Coverage of Antigravity CLI, Cursor ownership, Amp, the Anthropic third-party harness ban and its reversal, and Matt Pocock's and Karpathy-style skills collections did not surface in this lane's searches and remains a gap.