AutoHarness: open-source CLI makes agents split tasks, find skills, and face blind review
Professional_Hat9893 · reddit · 2026-10-21
The author noticed a common failure mode in agents: they skip the reference, work from memory, and grade their own output — even Opus 5.5 drawing a Blade Runner 2049 scene never looked at the actual frame. So instead of fixing outputs, he fixed the agent, turning each observed failure into a rule, resulting in AutoHarness (MIT, Python, no runtime deps):
- Areas of know-how: tasks are split into domains; each gets a skill from skills.sh or the MCP registry, or the agent writes one — methods must come with source links, not memory
- The original, not a description: real API responses, data samples, mockups, with facts measured before the first attempt
- A brief before any edit: quality bar, verbatim constraints, and how work will be checked; a hook keeps write tools closed until it's done
- Self-setup: the agent installs needed MCP servers, browser and skills after a trust score, scan and constraint check, removing only what it installed
- Author isn't the judge: a fresh subagent reviews item by item, with blind A/B against earlier results for visual work
- Learn before fixing: serious findings require skill changes before retry; three repeats trigger an approach change, an arbiter, then a human
- It accumulates: skills, lessons and setups carry over
Same model (Opus 5.5) with vs. without the harness: better results, biggest gains on visual work. Works with Codex, Gemini CLI, Claude Code, Antigravity, Qwen Code and Strands.
More from coding & agent
- Training an original-character LoRA for MiniMax H3: full workflow and pitfalls — DeliciousGorilla · 2026-10-21
- Dev builds open-source Chrome extension unifying Gemini and Antigravity on Windows — iamfivestars · 2026-10-21
- Dev interviews his Claude Code agents on a video standup — the human did worst — stickygoose · 2026-10-21
- MCP server SDK CVE-2026-111770: shared McpServer leaks responses across clients — Future_AGI · 2026-10-21
- When a Customer's AI Assistant Calls Support, Who Checks What It's Authorized to Do? — IrfanZahoor_950 · 2026-10-21
- Two weeks in SF: fundraising goes binary, top devs now review agents not code — ai · 2026-10-21