The archive · AI & Models · Product decision · 2026
AutoAgent bets agents can engineer their own harnesses; GitHub 4.6k stars in five months
AutoAgent automates harness engineering: a meta-agent edits the prompt, tools and orchestration, benchmarks each change, keeps score gains — 4.6k GitHub stars.
AutoAgent (Thirdlayer)
What the business is
AutoAgent is an open-source framework from Thirdlayer (thirdlayerinc) that automates agent-harness engineering: a meta-agent modifies the agent system prompt, tools, configuration and orchestration, runs Harbor benchmark tasks, checks the score, and keeps or discards each change, with a commercial product around self-configuring agents on a waitlist.
How it started
Thirdlayer (thirdlayerinc) published AutoAgent on GitHub on 2026-04-02, framing it as 'like autoresearch but for agent engineering': give an AI agent a task and let it build and iterate on its agent harness autonomously overnight. The founding premise stated in the README is that engineers should no longer touch harness Python files directly; the meta-agent modifies the system prompt, tools, configuration and orchestration, runs the benchmark, checks the score, and keeps or discards the change.
What happened
The repo shipped a deliberately small architecture — one editable agent.py with a fixed Harbor adapter, program.md as the human steering file, tasks in Harbor format — and the team opened a waitlist for 'a product around self-configuring agents' while hiring engineers through hello@thirdlayer.inc. Adoption followed quickly: by the 2026-09-05 crawl the repo showed 4.6k stars and 500 forks despite only two commits on main, and the README pointed contributors to Harbor task format and to context-engineering skill packs to lift agent performance.
No ending yet — it is still running.
Background
AutoAgent is an open-source framework by Thirdlayer (thirdlayerinc) for what its README calls autonomous harness engineering. Published on GitHub on 2026-04-02, it lets an AI agent build and iterate on its own agent harness overnight: the meta-agent modifies the system prompt, tools, agent configuration and orchestration, runs the benchmark, checks the score, and keeps or discards each change before repeating the loop.
The design moves the human out of the code. The harness under test lives in a single agent.py, the meta-agent instructions and build directive live in program.md, a Markdown file edited only by the human, and evaluation tasks follow Harbor format with each test suite producing a numeric score. The meta-agent hill-climbs on that score; Docker isolation contains autonomous edits; results accumulate in results.tsv.
The startup treats the repository as the wedge into a commercial product: the README banner advertises a product around self-configuring agents with a signup form and an open call for engineers. Adoption has been fast for an early repo — 4.6k stars and 500 forks by the 2026-09-05 crawl, about five months after creation — while main itself still carries only two commits.
What has to be true
- Measurable traction without press: a repository created 2026-04-02 reached 4,570 stars and 500 forks by the 2026-09-05 crawl, proof the agents-engineering-agents idea draws viewers on GitHub alone.
- A sharply stated bet: the README one-paragraph thesis — do not touch harness Python files, program a meta-agent instead — makes the wager falsifiable and easy to track over time.
- An explicit product path: the same page that explains the framework opens a waitlist for a commercial product around self-configuring agents and invites engineer applications.
- A clean mechanism: score-driven hill-climbing with Docker isolation is the same loop as autoresearch, applied to the harness itself, so progress stays auditable change by change.
What can be applied
Once a system is numerically scored, automate the improvement loop: let it edit its own configuration, keep only score-increasing changes, and the human job becomes the directive, not the code.
Aftermath
As of 2026-09-05 AutoAgent remains at the adoption-before-product stage: the open-source repository is live with 4,570 stars (4.6k), 500 forks, 29 watchers and two commits on main, while the README keeps advertising an upcoming commercial product around self-configuring agents, with a signup waitlist and open engineering hiring through hello@thirdlayer.inc. The material record shows no funding, revenue, pricing, customer or team-size disclosures and no third-party press coverage, so the visible assets are the repository, the waitlist and the hiring call.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card