AGENTS.md Template for Product Manager AI Delivery Agents

Overview

AGENTS.md template for PM AI delivery agents defines a structured operating context that supports both single-agent and multi-agent orchestration. It specifies roles, rules, and governance needed to ship AI-powered product features with accountability and observability. Direct answer: This template provides the project-wide operating context for PM-led AI delivery using AI coding agents and multi-agent collaboration.

It helps product teams codify the agent workflow, handoffs between planner, implementer, reviewer, tester, researcher, and domain specialist, and establishes a repeatable pattern for tool access, memory, and source-of-truth management.

When to Use This AGENTS.md Template

• Starting a new AI-enabled product initiative that requires coordinated agent work across planning, execution, and validation.
• Defining a reusable operating model for PM-driven AI delivery that scales to multiple features or products.
• Establishing governance around tool access, secrets, data sources, and production deployment for AI agents.
• Creating a handoff protocol between agents to minimize context loss and ensure quality assurance.
• Instituting a clear escalation path and human review where automated outputs require verification.

Copyable AGENTS.md Template

# AGENTS.md

Project: Product Manager AI Delivery

Goal: Enable PM-driven AI feature delivery using a defined roster of agents and an orchestrator with clear handoffs.

Agent roster and responsibilities
- Planner: defines product goals, AI features, user stories, acceptance criteria
- Implementer: builds agent behaviors, orchestrations, and tool integrations
- Researcher: sources data, validates facts and constraints
- Reviewer: ensures outputs meet policy, quality, and design standards
- Tester: executes tests, monitors interactions, verifies end-to-end flows
- Domain Specialist: provides domain constraints, risk assessment, and business context

Supervisor or orchestrator behavior
- Orchestrator coordinates all agents, enforces the plan, and maintains a single source of truth for the run
- Maintains memory of recent actions, decisions, and context within a defined window

Handoff rules between agents
- Planner completes the spec and passes to Implementer with context
- Implementer passes to Researcher for data and constraints as needed
- Researcher returns validated inputs to Implementer
- Implementer delivers to Reviewer for validation, then to Tester for execution
- If issues arise, hand back to Planner with rationale and proposed changes

Context, memory, and source-of-truth rules
- Central repository stores the product backlog, design docs, and requirements
- Memory is bounded (recent context window) to prevent drift and ensure relevance
- All outputs cite sources of truth and are traceable to the plan

Tool access and permission rules
- Agents access tools via a governed interface with scoped permissions
- Secrets are stored in a secure vault and never hard-coded
- Production actions require explicit approvals and audit trails

Architecture rules
- Event-driven, idempotent actions, single source of truth
- Clear boundaries between agents and shared services

File structure rules
- Repositories organized by agents, workflows, docs, and tests
- Naming consistent with the PM AI delivery pattern

Data, API, or integration rules when relevant
- Data sources are sanctioned and versioned
- API calls respect rate limits and privacy constraints
- All external integrations are logged and auditable

Validation rules
- Outputs link to acceptance criteria; automated checks exist where possible
- Outputs reviewed for policy and quality

Security rules
- Data privacy, least privilege, secrets management, and encryption in transit

Testing rules
- Unit, integration, and end-to-end tests for agent interactions
- Regression tests when workflows change

Deployment rules
- CI/CD with canary deployments and rollback plans
- Telemetry and health checks in production

Human review and escalation rules
- Any uncertain outputs require human review before production
- Escalations routed to PMs or security leads when risk is detected

Failure handling and rollback rules
- Roll back to the last known good state; preserve traceable artifacts
- Notify stakeholders and log the incident for postmortem

Things Agents must not do
- Do not access or reveal secrets
- Do not bypass approvals or production safeguards
- Do not modify production data directly without validation and audit
- Do not drift from the agreed plan or introduce unapproved changes

Recommended Agent Operating Model

The PM AI delivery operating model assigns clear boundaries and decision rights across the agent roster. The Planner creates and maintains the product vision and acceptance criteria; the Implementer translates the plan into agent rules and orchestrations; the Researcher validates data, sources, and constraints; the Reviewer validates outputs against policy and quality; the Tester performs end-to-end checks; the Domain Specialist ensures domain-specific correctness and mitigates risk. The Orchestrator enforces the plan, coordinates handoffs, and ensures traceability. Escalation paths exist for when outputs cannot be validated automatically, and human review becomes the gate for risky decisions or production changes.

Recommended Project Structure

pm-ai-delivery-project/
├── agents/
│   ├── planner/
│   │   └── plan.md
│   ├── implementer/
│   │   └── orchestrator.js
│   ├── researcher/
│   │   └── data-sources.md
│   ├── reviewer/
│   │   └── validation.md
│   ├── tester/
│   │   └── tests.md
│   └── domain-specialist/
│       └── domain-notes.md
├── workflows/
│   └── pm-ai-delivery/
│       ├──README.md
│       └── run.yaml
├── docs/
│   └── agent-handbooks.md
└── tests/
    └── end-to-end/
        └── e2e-spec.md

Core Operating Principles

• Clear ownership and accountability for each agent role
• Idempotent actions and deterministic outputs where possible
• Single source of truth and auditable decisions
• Data minimization, privacy by design, and secrets governance
• Explicit handoffs with traceable context and rationale
• Escalation to human review for uncertain or risky decisions

Agent Handoff and Collaboration Rules

• Planner to Implementer: hand off with complete spec, acceptance criteria, and reference to sources of truth
• Implementer to Researcher: request data, constraints, and validation inputs; return validated data
• Researcher to Implementer: provide structured inputs aligned to the plan
• Reviewer to Tester: deliver validation outcomes and pass/fail criteria
• Tester to Orchestrator: report test results and any failures or drift
• Domain Specialist to Planner: provide risk assessment and domain constraints; update backlog if needed

Tool Governance and Permission Rules

• All tool calls are scoped and auditable; no elevated privileges without approval
• Secrets are accessed via a secure vault; never hard-coded
• Production API calls require pre-approval and logging
• Data access is restricted to the minimum set needed to complete tasks
• Changes to production workflows must go through a review and deployment gate

Code Construction Rules

• Write modular, testable agent rules with clear interfaces
• Avoid global mutable state; prefer idempotent state transitions
• Document decisions and rationale in planning artifacts
• Do not copy-paste from unknown sources; cite references if used
• Use versioned data models and backward-compatible schemas

Security and Production Rules

• Encrypt sensitive data in transit and at rest
• Enforce least privilege across all agents and services
• Monitor for anomalous agent activity and have an incident runbook
• Require human review for production changes and data-altering actions

Testing Checklist

• Unit tests for each agent rule and function
• Integration tests for handoff paths and orchestrator behavior
• End-to-end tests covering planning, implementation, validation, and deployment
• Regression tests when workflows are updated

Common Mistakes to Avoid

• Skipping human review for risky decisions
• Allowing context drift across handoffs without a traceable rationale
• Bypassing the central source of truth or secret management
• Unclear ownership leading to duplicated work
• Unscoped tool access causing security risk

FAQ

What is this AGENTS.md template for?

This AGENTS.md template defines a PM AI delivery operating model and multi-agent orchestration pattern for product teams building AI features.

Who are the key roles in this template?

Planner, Implementer, Researcher, Reviewer, Tester, and Domain Specialist, with an Orchestrator coordinating handoffs and memory.

How are handoffs between agents managed?

Handoffs include explicit context, links to the source of truth, and acceptance criteria. Each transfer requires rationale and is auditable.

What governance exists for tool access and secrets?

Tool calls are scoped, secrets reside in a vault, and production actions require approvals and traceability.

How is success measured and validated?

Success is tied to acceptance criteria, policy compliance, automated validations, and human review when needed.