i.inc
A personal company of AI employees. I file tickets; agent employees work in teams, each in a walled sandbox, and hand back work that is checked and reviewed by a peer before it comes to me.
Coding agents are good at a single task. They are bad at being trusted with a repository while nobody is watching. Most of what it takes to trust one has little to do with the model: isolation, review, proof that the work is done, recovery when CI goes red, and memory that carries from one task to the next. i.inc is my attempt at all of that, and I use it every day.
How a ticket moves
An employee picks up a ticket on a fresh branch, plans, builds, and runs the project’s own checks, with a few fix attempts before it switches engines. A separate verifier then checks every “done when” item. Review happens with fresh eyes, on a different vendor’s model where one is available, for at most two rounds; anything still disputed goes into the report instead of looping forever.
Then the host, not the agent, pushes exactly the commit that was checked. If CI
on main goes red, the same ticket becomes the fix, and if it is still red two
hours later the push is reverted.
Three decisions
The team is the trust boundary. Agents need root to do ordinary work, so a Unix user per employee would isolate nothing. Each team gets one sandbox, the sandboxes cannot reach each other, and the real wall is at GitHub.
Agents never push. Sandboxes get a read-only token. A GitHub App holds the only write access, so a secret like a release signing key stays out of an agent’s reach because GitHub enforces it, not because the agent behaves.
Review on a different engine. A model reviewing its own diff mostly agrees with itself. Swapping engines is only possible because every engine plugs in through the same protocol.
Protocols
Every engine (Claude Code, Codex, Antigravity, or a local model through OpenCode) is an ACP agent, so an employee keeps its identity and memory when it moves to another vendor. The live side runs on Microsoft’s Agent Host Protocol: I can open any session from my phone, answer a permission card, or replay what an agent did at each stage. MCP tools are the next step. I did trust and lifecycle first, because reach was never the bottleneck.
Where it stands
About 150k lines of TypeScript (a Bun daemon, a Lit web app and a pure core), 117 pull requests in its first six weeks. Since a redesign in October the agents have written some of their own commits and filed their own follow-up tickets.