engineering-agents

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: