I use AI agents inside a strict process. I own the plan, the review and every line that ships. This page shows that process, at ShareTrip and on side projects.
AI agents
The loop
- PlanI approve it first
- BuildOne agent, one worktree
- ReviewA different provider
- FixOne round
- CheckReal tests, real results
- MergeI do this part
Every change follows the same path. The agents do the middle four steps. The first and the last are mine.
The rules
- No agent starts before I approve a plan. The plan says how to split the work and which model does each part.
- Each agent works alone. It gets its own worktree, so two agents never edit the same code.
- A different provider reviews the code. An agent that wrote a change tends to agree with it.
- Checks must really run. A report lists what passed, what failed, and what was skipped.
- Agents never merge. They do not push to shared branches. I integrate the work.
- Agents never touch production secrets. They inspect a live system only with my approval.
Threads
One long thread
Short threads
- handoff
- handoff
- handoff
- handoff
Skills
A skill is a written procedure that an agent loads for one kind of task. I write mine after I catch myself explaining the same thing twice. Every agent on every machine then follows the same steps.
- InvestigateExplain the code, the logs or the bug first. An investigation is not permission to fix.
- Inspect systemsRead databases, logs and storage safely: my approval first, careful credentials, careful data.
- Verify changeChoose checks that prove the change works when unit tests cannot.
- OrchestrateDecide when work is worth splitting, which model does each part, and how the pieces come back together safely.
- ReleasePromote changes across branches and repositories, then verify the deployment.
- HandoffPack the context of unfinished work so it continues in a new thread or on another machine.
What stays with me
- The requirements, and the question of what to build.
- The architecture and the trade-offs.
- The final read of every diff.
- The merge, and the result after it.
On my team sheet, the AI agents play false 9: they drop deep and set up the finish. The manager picks the team and owns the result.
Tools
Agents Claude Code · Codex
Workspace T3 Code, with one thread and one worktree for each task
Rules One config repository, shared by every machine
Secrets A password manager, and encrypted bundles between machines