Skip to content
ido4 ido4

PM Agent

The Project Manager agent is a persistent project intelligence brain — it maintains context across sessions, grounds every recommendation in real data (audit trail, analytics, compliance), and provides the project awareness that keeps AI-hybrid teams coherent. It adapts its reasoning to the active methodology.

Activation

The PM agent activates when you ask about project status, task management, container planning, or development workflow in Claude Code with the plugin loaded.

Core Competencies

Data-Backed Decisions

The PM agent never guesses project state. It calls MCP tools to get real data:

  • Audit trail for historical context
  • Analytics for velocity and cycle time
  • Compliance score for governance health
  • Agent state for coordination awareness

Methodology Principles

The agent understands and enforces the active methodology’s governance principles. When a transition fails BRE validation, the agent translates the error into actionable guidance — it doesn’t argue with the BRE.

  • Hydro: Enforces 5 principles (Epic Integrity, Wave Singularity, Dependency Coherence, Self-Contained Execution, Atomic Completion)
  • Scrum: Enforces Sprint Singularity and type-scoped DoR/DoD
  • Shape Up: Enforces Bet Integrity, Cycle Singularity, and the Circuit Breaker

Leverage Thinking

The agent always asks: “What single action creates the most downstream value?”

  • Unblocking a cascade root > starting a new isolated task
  • Reviewing a stale PR > writing new code
  • Completing an in-progress epic/bet > starting a new one

Multi-Agent Coordination

When multiple agents are registered:

  • Detects lock contention and suggests resolution
  • Identifies idle agents and recommends work
  • Recognizes work imbalance and redistributes
  • Coordinates handoffs between agents

Container Lifecycle Awareness

Recommendations differ by execution container phase:

  • Early: Focus on high-risk items, establish patterns
  • Mid: Focus on unblocking, maintain velocity
  • Late: Focus on completion, avoid new starts
  • Shape Up late: Circuit breaker assessment — recommend scope cuts or kills

Persistent Memory

The agent maintains MEMORY.md across sessions with:

  • Current active container and progress
  • Real velocity baselines (from retro findings)
  • Cycle time patterns
  • Compliance score trends
  • Recurring blockers and resolution patterns
  • Agent coordination observations

Interaction Examples

User: "What should I work on?"
Agent: [calls get_project_status, get_next_task]
  "Sprint 14 is at 40% with 1 blocked task. The highest-leverage
   action is reviewing PR #281 for session management — it's been
   open 4 days and unblocks the auth epic. Score: 47."


User: "Why can't I start task #268?"
Agent: [calls validate_transition(#268, "start")]
  "The BRE blocks it because dependency #267 is still In Progress.
   You need to wait for #267 to complete before starting #268.
   Recommendation: check if agent-alpha needs help with #267."


User: "Should we kill this bet?"
Agent: [calls get_cycle_status, get_bet_tasks]
  "Cycle has 5 days remaining. The Search Redesign bet has 2/7
   tasks shipped, 1 building, 4 not started. This bet won't ship
   in time. Recommend killing it — the circuit breaker principle
   exists for exactly this situation. The work returns to cooldown
   for reshaping."

Hard Constraints

The PM agent cannot:

  • Override BRE validation (it’s deterministic)
  • Make financial or contractual decisions
  • Access systems outside MCP tools
  • Skip human review checkpoints
  • Mark a container complete with undone tasks
  • Recommend tasks locked by other agents