Retrieve, grade the results, rewrite the query if they're weak, then answer. Stops answering from bad context.

Make a plan up front, run the steps one at a time against a repo, and replan when a step fails.

Draft, critique the draft, revise. Repeat until a critic signs off. The evaluator-optimizer pattern.

One supervisor routes each task to the right specialist (researcher, coder, writer) and stitches the results.

When something breaks, pulls the recent commits, guesses the likely cause, and posts a summary to Slack.

Reads merged PRs since the last tag, groups them by type, and opens a clean changelog PR.

Reads a proposed schema change, checks it against the live database, and flags the risky operations.

Searches your competitors, pulls features and pricing, and builds a comparison table you can actually use.

Plan sub-questions, search them in parallel, then synthesize one cited brief. Built for hard questions.

Answers questions about a codebase with citations to the exact file, and says so when the answer isn't there.

Queries your metrics tables on a schedule, spots anomalies, explains them, and pings Slack with context.

Reads yesterday's Jira and GitHub activity and posts a tight standup summary to Slack every morning.