All work
Proof of conceptsince October 2026
G7 Agentic Platform
A prototype for taking a small code change through testing and review.
I designed and tested the prototype.
Why I built it
I wanted to see what it would take to review a change written by an agent and undo it if a problem showed up after release.
How it fits together
- Requestplain English
- Brief agentquestions, then a spec
- Builder agentedits allowed paths only
- Six QA agentson staging
- Release gatethree people approve
- Monitorrolls back on failure
Implementation
- A brief agent that asks before it builds
- It asks at most three questions, each with a recommended answer, then writes a spec with acceptance criteria.
- A builder with a narrow scope
- The builder agent (Claude Opus) edits a working copy of the production commit. It can write only to paths marked editable, and every write is validated.
- Six QA agents on staging
- They review the change in parallel, and only issues the change introduces count. Blocking issues go back to the builder once; anything still blocking stops the release. A check that times out or returns nothing readable counts as a failure, never a pass.
- Release approval
- Three approvals from three different people in Product, Engineering and QA, excluding the author. The exact commit that was staged is the one that ships.
- Automatic rollback
- After a release it checks every 30 seconds for ten minutes, then every five minutes. Two failures caused by the release roll it back automatically.
- Least privilege
- The platform holds no hosting API key. It deploys with a repository-scoped key and signed webhooks, and app code runs on a separate runner that has no secrets.
Results
- to roll back a bad release automatically
- 63 sto roll back a bad release automaticallylive fault drill
- from prompt to production
- 5 minfrom prompt to productiontest run, $2.56 of model usage
Stack
- TypeScript
- PostgreSQL
- Claude Opus via the Anthropic API
- Agent tool use
- git
- Next.js 16
- Hosted on G7Cloud