Agents for engineering teams

Agents that give your dev team its time back.

These are the agents I build for teams that want their engineers writing code, not doing the work around the code. Each one replaces a specific piece of busywork, not the engineer.

Every one of them runs on the same architecture as the toolkit: state graph, tools over MCP, human approval before any irreversible action.

01

Code review

What it does

Reads every pull request, flags obvious bugs, suggests improvements, checks the diff against the team's style guide. Adds inline comments like a human reviewer would.

When it's worth it

When code review is the bottleneck slowing releases down. Small teams where senior engineers spend half their day reviewing juniors.

Where it fails

Should not block merges on its own. Its comments are suggestions, not gates. A human still approves.

02

Bug triage

What it does

Reads new issues from GitHub, Linear or Jira. Labels by area, assigns severity, checks for duplicates, pings the right person on Slack.

When it's worth it

Any team drowning in issue backlog. The agent turns a 200-issue queue into a triaged, sorted one overnight.

Where it fails

Wrong severity on a real incident is worse than no triage. Confidence threshold: if the agent isn't sure, it flags for human review instead of guessing.

03

Dependency updates

What it does

Watches package.json, requirements.txt, Cargo.toml. Opens PRs for security patches. Runs the test suite. Summarizes what changed and what broke.

When it's worth it

Dependabot exists but is dumb. This one reads changelogs, groups related updates, and doesn't spam PRs.

Where it fails

Never auto-merges. Every PR needs human approval, because a supply-chain attack via a bad dependency is the exact scenario this agent could accelerate if trusted blindly.

04

On-call triage

What it does

First responder for alerts. Reads the alert, checks recent deploys, looks at logs, tries known runbooks. Pages a human only when it can't resolve or when the alert matches a serious pattern.

When it's worth it

Small teams without a dedicated on-call rotation. Cuts pages by 60-80% for teams with noisy alerts.

Where it fails

Silencing a real incident is the failure mode. Every silenced alert gets a human review the next morning, no exceptions.

05

PR summaries

What it does

Reads a large PR and writes a plain-language summary at the top: what changed, why, what to test, what to watch out for in review.

When it's worth it

Teams where PRs regularly hit 500+ line diffs and reviewers skim instead of reviewing. Saves hours per week per reviewer.

Where it fails

Low. The worst case is a bad summary, and the reviewer still reads the code.

06

Docs sync

What it does

Watches the codebase. When public API changes, updates the docs. When examples break, flags them. When a new feature ships without docs, opens a doc PR.

When it's worth it

Teams where docs are always out of date because updating docs is the last thing anyone wants to do.

Where it fails

Auto-generated docs read like auto-generated docs. Agent writes a draft, human polishes. Never publishes without a human pass.