Products · Sigma Loop · Open source
AI coding agents that plan, get reviewed and leave a record.
Sigma Loop is an open-source tool that makes AI coding agents plan, get reviewed and leave a record. It is the set of guardrails we build our own products with: nothing is edited until a plan has been challenged, and nothing is called done until it has been checked.
1
Plan
The agent writes a plan before it changes anything.
Left alone, a coding agent goes straight to editing. Sigma Loop puts every change that is not trivial through seven phases, in order.
Goal. What is wanted, and how anyone will know it is done.
Research. What the code does today, and what the change will touch.
Plan. The steps, the files and the tests, written down.
Plan review. The plan is challenged against the real code.
Implement. The work, test first.
Review. The finished change is reviewed.
Retrospective. What was learned is kept for next time.
2
Review
The plan is challenged before any edit. The result is reviewed before it counts.
A separate review reads the plan against the code as it really is, and against your product strategy if you have written one down. A second review reads the finished change. Each can send the work back.
Careful is faster.
A wrong turn caught in the plan costs minutes. Caught after the code is written, it costs the code.
3
Check
Done is checked, not claimed.
A goal is not marked done until the record shows that every phase ran and both reviews approved. Once you have named your test command, the tests must pass too.
Some of these rules are enforced by code and some are instructions to the agent. The documentation says which is which.
4
Record
Every phase leaves a record.
Each phase is written to the ticket it belongs to, or to a journal kept with the code, with its cost where the AI provider reports one. Months later you can read what was planned, what review found, and what changed.
Work can be queued and run through in order, one change request per goal. Anything that needs a person is set aside with the reason, not guessed.
AI does the work. People make the decisions.
The agent plans, builds and checks. When a decision belongs to a person, it stops and says why.
5
A replay
One real piece of work, start to finish.
Goal 2135 in Sigma Loop's own repository, run by Sigma Loop. Step through it: the goal, the plan, the review, the finished change.
1. Goal
Goal 2135
2. Plan
Three of its decisions
3. Review
What review found in the first version
MESSAGE = ("Commit signing is unavailable on this machine, so nothing can be committed or pushed. " "Start a fresh session (in a Claude Code cloud container the signer comes back when the " "container restarts). Sigma's commit and publish steps will not turn signing off for you.") # …PARK_TAIL = "Do not retry: record this goal as parked with this message."From the public repository. The change touched 15 files, three of them new test files.
6
Measured
What has been measured is published with the code.
The repository holds a page of measured results: cost per goal and per phase, review findings, test results. Each figure names its source file and its date. It makes no comparison with any other tool.
7
Open source
Free to use, and the code is public.
License. MIT.
Works with. Claude Code, fully validated. Codex, partially validated. Cursor, experimental.
Needs. git, bash and Python 3.10 or newer. Nothing else.
8
In your team
We can set it up inside your engineering team.
We install Sigma Loop with your engineers, write down with you what a person must approve before work moves on, run a first backlog together, and hand it over.

