AgentEM - Engineering Management Agents for Claude Code

6 AI agents that handle specs, ticket decomposition, risk detection, review routing, release management, and sprint retrospectives. They remember what they've done, take real actions, chain intelligently, and learn from your feedback. Encoded with your team's engineering judgment.

Install

/plugin marketplace add anicol/engineering-agents
/plugin install agentem@engineering-agents

What you get

Command	What it does
/agentem:spec-generator	Turns product briefs into specs, saves them, creates tracking issues
/agentem:ticket-decomposer	Breaks specs into tickets, creates GitHub issues in batch or individually
/agentem:risk-detector	Scans for risks, creates issues for critical risks, comments on stale PRs
/agentem:review-orchestrator	Assigns reviewers to PRs, nudges stale reviews, balances review load
/agentem:release-manager	Generates release artifacts, creates GitHub releases, saves changelogs
/agentem:retro-analyzer	Generates retro docs, updates learnings, creates action item issues

Quick start

1. Install the plugin and open your project in Claude Code.

2. Scaffold context files:

   /agentem:init
   

   This creates 8 context files in context/ - your product strategy, architecture map, team topology, capacity, review standards, spec conventions, and learnings (what works + what doesn't).

3. Fill in context files. Start with context/product/strategy.md. The more specific you are, the better the agents perform. "We use microservices" is useless. "We have 4 services: auth (Python/FastAPI), api (TypeScript/Express), worker (Go), dashboard (Next.js)" is useful.

4. Check your progress:

   /agentem:doctor
   

5. Run an agent:

   /agentem:risk-detector
   /agentem:spec-generator
   

6. Run the full flow:

   /agentem:sprint-plan Build user notification preferences
   

   This chains spec generation, ticket decomposition, and risk detection.

All commands

Command	What it does
/agentem:spec-generator	Turns product briefs into specs, saves them, creates tracking issues
/agentem:ticket-decomposer	Breaks specs into tickets, creates GitHub issues in batch or individually
/agentem:risk-detector	Scans for risks, creates issues for critical risks, comments on stale PRs
/agentem:review-orchestrator	Assigns reviewers to PRs, nudges stale reviews, balances review load
/agentem:release-manager	Generates release artifacts, creates GitHub releases, saves changelogs
/agentem:retro-analyzer	Generates retro docs, updates learnings, creates action item issues
/agentem:sprint-plan	End-to-end: spec → tickets → risk scan
/agentem:init	Scaffold context/ directory with 8 template files + autonomy config
/agentem:doctor	Check context files, autonomy config, agent state, and agent readiness
/agentem:status	Dashboard: agent activity, risks, PR status, sprint health, effectiveness
/agentem:watch	Poll GitHub for events and trigger agents (new PRs, stale PRs, merges)

Context files

The agents are only as good as the context you give them. Generic prompts produce generic output.

File	What it does	Who uses it
product/strategy.md	Mission, OKRs, priorities, what you're NOT doing	Spec Generator, Risk Detector
architecture/system-map.md	Services, ownership, data flows, constraints	Spec Generator, Risk Detector, Review Orchestrator
team/topology.md	People, roles, ownership map, skill matrix	Ticket Decomposer, Review Orchestrator, Risk Detector
team/capacity.md	Sprint capacity, estimation approach, planning rules	Ticket Decomposer
standards/review-playbook.md	Review philosophy, SLAs, focus areas, patterns	Review Orchestrator
standards/spec-standards.md	Spec structure, conventions, quality bar	Spec Generator
learnings/what-doesnt.md	Anti-patterns to avoid (updated after retros)	All agents
autonomy.yaml	What agents can do without asking	All agents

Autonomy config

By default, agents ask before taking any action. Edit context/autonomy.yaml to control this:

# Actions agents execute without asking
autonomous:
  risk-detector:
    - scan

# Actions that require your approval (default)
requires_approval:
  review-orchestrator:
    - assign-reviewers
  spec-generator:
    - save-spec

# Actions agents will never take
disabled: {}

Agents also remember what they've done between runs, learn from your feedback, and offer to chain to relevant follow-up agents when they finish.

Security & permissions

Agents use the gh CLI for all GitHub interactions. Here's what they can do, grouped by risk:

Risk tier	Actions	Agents	gh command
Read-only	Scan repos, PRs, issues	Risk Detector	gh pr list, gh issue list
Write (local)	Save specs, retros, release notes to your repo	Spec Generator, Release Manager, Retro Analyzer	File writes only
Write (GitHub)	Create issues, comment on PRs	All 6 agents	gh issue create, gh pr comment
Write (assignments)	Assign PR reviewers, create releases	Review Orchestrator, Release Manager	gh pr edit --add-reviewer, gh release create

Agents never perform destructive actions. They will not close issues, merge PRs, delete branches, force-push, or modify sprint scope. These are outside agent scope by design.

Every action above defaults to requires_approval — the agent shows you the exact command and asks before executing. You control this per-action in context/autonomy.yaml. The template includes risk tier annotations so you can make informed decisions about what to make autonomous.

What this doesn't include

• Background automation - Watch mode runs in-session. No cron scheduling, webhook triggers, or Slack delivery.
• Integrations - No Linear/Jira/Slack API connections. Agents use gh CLI when available, filesystem otherwise.

These are part of the paid consulting tier. The plugin gives you the agents, actions, memory, and context structure. The consulting engagement adds extraction interviews, background automation, integrations, and tuning.

Ready to go further?

Most teams get the plugin running in 30 minutes, then stall on context because extracting engineering judgment is a different skill than prompting.

Book a 30-minute call - we'll map your workflow and show you what the system looks like with your team's actual context.

Staying up to date

Third-party plugins don't auto-update by default. To get new versions automatically, run:

/plugin marketplace update

This takes you to the marketplace manager where you can enable auto-update for engineering-agents. Once enabled, new versions are pulled automatically at session start.

Development

git clone https://github.com/anicol/engineering-agents.git
claude --plugin-dir ./engineering-agents/agentem

---

Built by AgentEM