ai agents automation claude-curated
Notes on getting real work out of AI coding agents — the practices that separate “impressive demo” from “reliable tool you lean on daily”.
Context is the scarce resource
An agent is only as good as what’s in its window. Practical moves:
- Give it a map, not the territory — point at the few files that matter, not the whole repo. Let it search for the rest.
- Persist durable facts outside the session (a memory file, a project doc) so you’re not re-explaining the setup every time.
- Delegate reading — have a sub-agent sweep many files and return only the conclusion, keeping the main context clean.
Tools and permissions
- Least privilege by default — allowlist the safe, high-frequency actions; deny the destructive ones (
rm, force-push, writes to secrets). The agent moves fast on the safe set and stops at the dangerous edge. - Read/write asymmetry — reading is cheap and reversible; writing and outward-facing actions deserve a confirmation gate.
Hooks and automation
The highest-leverage pattern is deterministic automation around a non-deterministic core: use lifecycle hooks (on submit, on tool use, on stop) to log, validate, or gate — things the model shouldn’t be trusted to remember to do every time. “Run the linter after every edit” belongs in a hook, not a hopeful instruction.
Parallelism and verification
- Fan out independent work — multiple agents on independent files/questions finish in wall-clock time, not sum-of-parts.
- Verify adversarially — for anything that matters, have a second pass try to refute the first agent’s result rather than confirm it. Plausible-but-wrong is the failure mode to design against.
- A build/test is the ground truth — an agent that says “done” without a green check hasn’t earned the claim.
Guardrails that matter
- Keep a human gate on irreversible or outward-facing steps (publishing, deleting, sending).
- Treat generated output as a draft to review, especially anything published under your name.
- Log what the agent did — reproducibility and auditability beat trusting memory.
Related: Building an MCP Server.