TAPE 04 / 07SELECTED WORK
THE CREDITS
Agentic Factory
I designed the trust model that lets AI agents take a well-scoped ticket from backlog to a reviewed pull request: no write credentials, a command guard that doesn’t rely on the model behaving, and evidence a reviewer can check. Only people can merge.
- BRAND
- Adobe
- INDUSTRY
- Developer Platform
- MY ROLE
- Architecture / Safety Design / Platform Engineering / Incident Response
- ENGINEERING
- GitHub ActionsGitHub Agentic WorkflowsClaudeOpenAI CodexAWS BedrockNode.jsJira APISlackzizmor / actionlint / poutine
- DESIGN
- Systems & safety design
THE CHALLENGE
A bug fix or a small feature still takes a senior engineer to read, code, test and review. The question was never whether an LLM could write the code. It was whether we could make it trustworthy enough to try in a real codebase.
THE APPROACH
- Safety by structure. Agents never hold write credentials. Pushes and PRs happen in a separate gated job, behind a command guard with no model in the loop.
- Policy per repo. Each repository declares its editable paths and prerequisites in a policy file, so onboarding is a config change.
- Evidence. Every run produces schema-validated code, security and acceptance reviews and test results, rendered into the PR.
- Blind spots. Checks the sandbox can’t run become visible, required CI follow-ups instead of being skipped.
- Saying “I don’t know.” When a decision belongs to a person, the agent posts its reasoning to the ticket and stops.
- Documentation and debugging. I maintained the architecture docs and debugged production gaps the platform’s own PRs surfaced.
THE OUTCOME
Well-scoped tickets now run end to end on their own, behind a mandatory two-approval human merge gate. Velocity went up and review rigor didn’t go down.