My agent setup is a decision system.

01

Start from the real state.

An agent entering a repo should recover what actually happened before it moves. The handoff, branch, transcript, tests, and current files all matter because summaries drift.

agent-session-resume ↗ starts from the handoff and surrounding evidence across Codex, Claude Code, Antigravity, and OpenCode. AgentCanvas gives that recovery work a local evidence surface, so the repo becomes easier to inspect with a person.

02

Keep judgment in skills.

A prompt handles the moment. A skill keeps the judgment that should survive it: where to look first, what proof counts, what to avoid, and when to ask for another view.

I keep source skills in their own repositories and symlink them into a shared skills directory. That lets the same operating knowledge show up across tools without copying fragile config around.

03

Route models by job.

Different models are useful for different calls. Codex is where I investigate, implement, and verify. Claude helps with taste, framing, and second-order risk. Gemini helps compress sprawling evidence. Grok helps pressure-test direction and scope.

The model-orchestrator ↗ makes the jobs explicit: decide, review, synthesize, ship. One person still owns the call. The system is there to make the call better informed.

04

Make access explicit.

Agents get more useful when they can reach repositories, browsers, deployments, data, and other tools. That access needs a control plane.

The MCP registry records what a capability does, which harness can use it, whether the adapter is proven, and when delegation needs approval. The agent should start from the catalog, route through an approved adapter, and keep evidence, execution, and access as separate concerns.

05

Run the loop.

1. Recover. Read the handoff, inspect the branch, and verify the current state before treating a summary as truth.
2. Decide. Surface the real call and make the tradeoff visible.
3. Build. Make the smallest coherent change that can solve the problem.
4. Review. Bring in fresh eyes when scope, correctness, or taste can change the outcome.
5. Prove. Open the route, hit the endpoint, inspect the deploy, or run the real workflow.
6. Continue. Write down what changed, what was verified, what remains uncertain, and the exact next action.

06

Audit the system itself.

Once a week I audit the work: transcripts, repeated failures, useful workarounds, and places where an agent needed too much supervision.

Each pattern gets classified. It might need no action, a memory note, an updated skill, a new private skill, a public skill, or an automation. The point is to notice where the setup is brittle and turn earned lessons into reusable structure.