Common mistakes

1 min read · Chapter 9 of 10

What usually goes wrong

Symptom	Likely cause	First thing to check
Agent pull requests pile up unreviewed	More output than review capacity; diffs too big	Review queue age; a size limit; route agents to verifiable work
The agent can't get started	The repo doesn't build in the sandbox: private registry, missing secret, a service it can't reach	The setup script, run from scratch
Tests pass but the change is wrong	Weak or flaky tests; the agent edited the tests	Test diffs in review; protected tests; code owners on test folders
The bill spikes mid-month	A few heavy users, parallel agents, cache misses	Usage by user and model; budgets and caps
Agents stop working for everyone	Pooled credits used up with overage off	The pool's burn rate in the first week
Production data gone	An agent with real credentials outside a sandbox	Who holds production credentials; backups outside the blast radius
A token leaks from CI	An AI agent in a workflow reads untrusted issue text and holds a powerful token	Who can trigger the workflow; the token's scope; untrusted fields in prompts
A dependency nobody chose	A hallucinated or brand-new package	Lockfile changes in review; minimum release age; registry allowlist
An auditor asks who approved a change	The prompter approved the agent's change	Branch protection; reviewer independence on regulated code