Build a coding-agent harness that rewrites itself from its failures
Based on a 2026 Harness Learning paper, a developer built a coding-agent harness that studies its own failed runs and rewrites itself. He kept the model (Claude Sonnet 5.5) and tasks fixed and only let the harness change.
Source
@singh_builds on X · 2026-10-17
Written in our own words. Credit to the original author. Link checked 2026-10-22.
What they did
Based on a 2026 Harness Learning paper, a developer built a coding-agent harness that studies its own failed runs and rewrites itself. He kept the model (Claude Sonnet 5.5) and tasks fixed and only let the harness change.
Result (as claimed)
On Terminal-Bench 4.0 tasks it had never seen, success rose from 50% to 67%.
Tools they used
Cost: Claude Sonnet 5.5: from $2 in / $10 out per 1M tokens.
Set it up
Pick a contained repo or task
Start on a branch or side project where mistakes are cheap.
Install and configure
Set up Claude Sonnet 5.5, Python. Read what it can run on your machine and limit permissions where possible.
Run it on a real task
Give it a well-scoped issue and let it work. Review the diff like any pull request.
Add guardrails
Keep tests and CI in the loop, and never let an agent push to main or touch production secrets unattended.
Copy the full setup prompt
Paste into your agent and it will walk you through the setup, asking before anything that sends, posts or spends.
You are helping me set up this AI use case end to end in Claude Code. Goal: Build a coding-agent harness that rewrites itself from its failures What the original author did: Based on a 2026 Harness Learning paper, a developer built a coding-agent harness that studies its own failed runs and rewrites itself. He kept the model (Claude Sonnet 5.5) and tasks fixed and only let the harness change. Reported result (as claimed by the author): On Terminal-Bench 4.0 tasks it had never seen, success rose from 50% to 67%. Source: https://x.com/singh_builds/status/2111391498310717782 Tools they used: Claude Sonnet 5.5, Python Category: Dev & ops Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
You are helping me set up this AI use case end to end in Codex. Goal: Build a coding-agent harness that rewrites itself from its failures What the original author did: Based on a 2026 Harness Learning paper, a developer built a coding-agent harness that studies its own failed runs and rewrites itself. He kept the model (Claude Sonnet 5.5) and tasks fixed and only let the harness change. Reported result (as claimed by the author): On Terminal-Bench 4.0 tasks it had never seen, success rose from 50% to 67%. Source: https://x.com/singh_builds/status/2111391498310717782 Tools they used: Claude Sonnet 5.5, Python Category: Dev & ops Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
You are helping me set up this AI use case end to end in Grok Bot. Goal: Build a coding-agent harness that rewrites itself from its failures What the original author did: Based on a 2026 Harness Learning paper, a developer built a coding-agent harness that studies its own failed runs and rewrites itself. He kept the model (Claude Sonnet 5.5) and tasks fixed and only let the harness change. Reported result (as claimed by the author): On Terminal-Bench 4.0 tasks it had never seen, success rose from 50% to 67%. Source: https://x.com/singh_builds/status/2111391498310717782 Tools they used: Claude Sonnet 5.5, Python Category: Dev & ops Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
The Daily 5
Get five new use cases tomorrow morning
Free. Real sources. Unsubscribe any time.
Guides for this use case
Related use cases
Benchmark your agent harness against Claude Code with hidden judges
The founders of Relay, a local-first agent supervisor, pre-registered a 5-test suite with hidden judges and ran their tool, Claude Code, Prime Agent and Hermes through the same token meter on Sonnet 5.5 and Haiku 5.5, then published the losses along with the wins.
Self-host an agent that builds mini apps from chat and edits its own UI
mobius from mobius-os is a self-hosted personal AI system: you chat with a coding agent through Claude Code or Codex and it builds mini apps on your server and modifies its own UI.
Call any coding agent harness by name from one sandboxed server
He built an open-source server where you name the harness (Claude Code, Codex, Pi or about 20 more) and send a request, with each run sandboxed. Use it for one-off jobs, build agents on top, or write loops instead of prompts.
Pack a coding agent harness into one MCP server you can drive from ChatGPT
A developer built CloudHarness MCP, a whole coding agent harness inside one MCP server: pi agents run inside it and carry out tasks delegated from ChatGPT, skills install from any GitHub repo, and credentials are injected without showing up in chat.