Skip to content
← Work

Systems · Working method

How I run an engineering practice with AI agents

I use agents to implement, review and translate bounded pieces of work. I own the scope, read the evidence, resolve conflicts and decide what is ready to release.

Status
In use in my engineering practice
My role
Workflow designer and orchestrator

Agent fleet

Separate roles, with a written handoff.

My fleet kit gives implementers, reviewers and translators different briefs. The reviewer checks the actual change and reports findings with file references. A report is something to inspect, not permission to release.

I use Claude Code for orchestration and Codex for bounded engineering tasks. The kit also supports OpenCode chains. I choose the role and scope before choosing the model.

Delivery chain

Implement. Review. Integrate.

  1. Implement in a worktree

    The agent owns a branch and a bounded task. Other agents work in separate worktrees. Changes stay inspectable as a diff.

  2. Read the verdict

    Report and review files carry the checks, findings and a PASS or FAIL verdict. I resolve real findings and explain anything outside the agreed scope.

  3. Check the integrated result

    I resolve conflicts, build the combined tree and run its gates. Release is a separate decision after those checks.

Sandboxes and data

A worktree separates changes. Permissions bound actions.

A separate checkout avoids agents editing the same files. It does not isolate credentials or grant access to production. I set the filesystem, connector and network permissions for the task separately.

For this site, agents prepare local branches and review packets. They cannot deploy, push, submit live forms or change measurement settings. Client material stays outside the public repository; examples and fixtures are anonymised or synthetic.

Review files

Evidence I can follow.

Report
What changed, what ran, what failed and what remains unverified.
Verdict
PASS or FAIL with specific findings and file references.
Browser captures
The actual built page at desktop and phone widths.

Review gates

Different checks answer different questions.

  • Code and policy

    Type checks, tests and route ownership catch broken contracts before integration.

  • Copy and data

    Claims checks and the client-data scanner guard the public output. New page wording still needs my approval.

  • Rendered result

    A successful build is followed by screenshots, links and responsive checks on the actual output.