How to coordinate a human-AI dev team
CodevOS combined one human, parallel implementation agents, versioned specifications, enforced workflow gates and three-model review to report 106 pull requests in 14 days.
news.ycombinator.comSuggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Coordinate [development initiative] using version-controlled specifications and plans. Split agreed implementation tasks among Builders in isolated git worktrees. Enforce [human gates], build verification and independent Claude, Gemini and Codex reviews before presenting pull requests for integration.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- waleedk
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reported 106 pull requests shipped in 14 days by one human with implementation agents, with higher cost and longer execution than an unstructured comparison.
- Human involvement
- The human writes and reviews planning documents, reviews pull requests and participates in enforced human gates.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Write and version specifications, plans and reviews alongside source code.
- Have an Architect agent spawn Builders in isolated git worktrees.
- Run implementation through Porch's human gates and build-verification loops.
The CodevOS author reported shipping 106 pull requests in 14 days as one human working with AI agents that performed implementation. Their system treats development as a joint human-agent workflow, with the human primarily writing and reviewing specifications, plans and other natural-language documents.
An Architect agent spawns Builder agents that work simultaneously in isolated git worktrees. Specifications, plans, reviews and agent instructions are version-controlled alongside source code. A state machine called Porch enforces human gates, build-and-verify loops and mandatory review phases, while preserving checkpoints for replacement agents when context runs out.
The author described independent Claude, Gemini and Codex reviews at every phase, reporting that the models caught different problems and no single model found more than 55% of the bugs. They also pointed to a controlled comparison against unstructured Claude Code, describing higher cost and longer execution alongside more bugs caught and tests shipped. Those comparison details were summarized rather than reproduced in the post.
Who it's for: developers prepared to direct parallel agents through written specifications, enforced process stages and repeated review.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · waleedk
Frequently asked questions
How were simultaneous implementation tasks separated?
Builder agents worked in isolated git worktrees, with branches supporting workflow and pull requests handling integration.
What tradeoff did the author report?
The author described CodevOS as costing more and taking longer than unstructured Claude Code, while catching more bugs and shipping with tests.
More details
Remaining steps
- Obtain independent reviews from Claude, Gemini and Codex.
- Integrate reviewed work through pull requests and merges.
About this account
Builder agents implement parallel tasks while Porch enforces process stages and an Architect agent coordinates work.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=47167466. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: waleedk on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
CodevOS · Porch · Git · Claude · Gemini · Codex
Agent orchestration · Code review · Git worktrees
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.