Anthropic’s Lydia Hallie describes her team’s evolving Claude Code workflow in a recorded AI Engineer talk.
Instructions and permissions
- Keep project instructions focused on context the agent cannot infer from the code; audit stale instructions as models change.
- Auto mode classifies commands, but Hallie says explicit deny rules are needed for actions that must not occur.
- She cites an Anthropic study in which auto mode blocked 89% of dangerous commands; the comparison is vendor-reported, and some still passed.
Boundaries need layers
- A permitted test can run untrusted dependencies, so command approval alone cannot control all child-process effects.
- Sandbox file and network access; check its coverage and protect secrets explicitly. Hooks, MCP servers and other processes may sit outside the shell sandbox.
- Cloud sessions isolate a fresh checkout, but pushed commits and connector messages still have external effects. Keep deny rules and review accordingly.
Verification at scale
- Define observable completion criteria and inspect actual test or build output; use hooks for checks that must run.
- Hallie says blocking hooks must exit with code 2, rather than a typical test’s exit code 1; validate behavior against the installed version.
- CI applies common checks to every pull request. Anthropic uses a separate agent review, with human review scaled to change risk.
- An orchestrator can split work across sessions, but reliable tests, boundaries and review remain prerequisites rather than automatic guarantees.
“This machine is disposable, but the things that leave it aren’t.”