Resources for structuring and operating a software delivery team that uses agentic AI development tools (GitHub Copilot, Cursor, Kiro, Claude Code, Devin, etc.).
├── team-structure.md # Research report: how traditional roles evolve with agentic AI tools
├── multi-repo-documentation.md # Research report: documenting multi-repo systems for agents
├── agents/ # Assisting agents for each team role
│ ├── ai-architect/ # Senior AI Architect / Tech Lead
│ ├── senior-engineer/ # Senior Engineer
│ ├── quality-engineer/ # Quality Engineer / SDET
│ ├── ba-product-analyst/ # BA / Product Analyst
│ ├── devops-platform/ # DevOps / Platform Engineer
│ ├── ai-execution-agent/ # AI coding agent (execution layer)
│ └── self-improving-agent/ # Agent that learns from session history
└── skills/ # Reusable agent skills (install into your AI tool)
├── deep-research/ # Rigorous, citation-backed research on any topic
├── create-agent/ # Design and configure a new AI agent
└── create-agent-skill/ # Author new agent skills (SKILL.md format)
A research-backed report (25 sources, May 2026) covering:
- How each traditional role changes when the team adopts agentic AI tools
- What gets automated, what becomes more important, what new skills are required
- Recommended pod structure and team topology
- Headcount implications and real company examples (Salesforce, Klarna, Coinbase)
Use this when planning a team transition, hiring, or explaining role changes to stakeholders.
A research-backed report (20 sources, May 2026) covering:
- The three-tier memory model (hot/warm/cold) for agent context
- Documentation architecture across 4 levels: system map → repo AGENTS.md → module docs → cold memory retrieval
- Cross-repo coordination patterns: shared contract repo, one agent per repo, hierarchical AGENTS.md
- Staleness prevention: PR checklists, CODEOWNERS, CI drift detector, auto-generation
- The 7 failure modes poor documentation causes and how to prevent each
Use this when setting up documentation for a multi-repo system that AI agents will work on.
Each subdirectory contains a ready-to-use agent configuration for one team role.
agents/<role>/
├── AGENT.md # Steering file — always-on context, rules, and constraints
└── skills/
└── <skill>.md # On-demand procedures, loaded when the task matches
AGENT.md is the agent's steering file: facts about the role, universal rules, and standing constraints. Keep it loaded at all times for that agent.
skills/*.md are on-demand procedures: step-by-step workflows the agent follows for specific tasks. Load a skill when the task matches its description.
| Agent | AGENT.md covers | Skills |
|---|---|---|
ai-architect |
Governance rules, risk tiers, SDLC pipeline, AI pattern vocabulary | agent-job-spec, ai-code-review, architecture-decision-record, tech-debt-triage, governance-checklist, incident-postmortem, ai-pattern-catalog |
senior-engineer |
Orchestration rules, blast radius guide, spec-driven development | spec-generator, agent-failure-diagnostician, context-file-curator, codebase-health-audit, pr-review |
quality-engineer |
Risk zones, six-layer testing strategy | adversarial-code-review, eval-suite-builder, test-strategy-charter, bug-report, regression-test-generator |
ba-product-analyst |
Spec depth guide, 7-section spec structure | spec-drafter, spec-completeness-auditor, impact-assessor, requirement-triage, acceptance-criteria-generator, stakeholder-communicator |
devops-platform |
Model routing tiers, cost control levers | agent-identity-governance, agent-cost-investigator, agentic-pipeline-architect |
ai-execution-agent |
Capabilities, guardrails, escalation triggers, PR template | task-kickoff, context-reader, pr-writer, escalation-handler, prompt-engineering |
self-improving-agent |
Feedback signal taxonomy, lesson quality rules, conflict resolution policy | learning-session |
Reusable agent skills used to build and maintain this repository. Install them into your AI tool to use them in your own projects.
| Skill | What it does |
|---|---|
deep-research |
Conduct rigorous, citation-backed research on any topic using iterative search and evidence grading |
create-agent |
Design and configure a new AI agent — steering file, skills, and the balance between them |
create-agent-skill |
Author new agent skills conforming to the agentskills.io open standard (SKILL.md format) |
Kiro — copy to your skills directory:
# User-scoped (available in all projects)
cp -r skills/deep-research skills/create-agent skills/create-agent-skill ~/.kiro/skills/
# Project-scoped (available in this repo only)
cp -r skills/deep-research skills/create-agent skills/create-agent-skill .kiro/skills/Claude Code — copy to your Claude skills directory:
cp -r skills/deep-research skills/create-agent skills/create-agent-skill ~/.claude/skills/Cursor / other tools — paste the relevant SKILL.md content into the conversation when starting a task that matches the skill's description.
- Copy the relevant
AGENT.mdcontent into your agent's system prompt or steering file - Copy the
agents/<role>/skills/files into your agent's skills directory - The agent will load skills automatically when the task matches the skill's description
- Paste the
AGENT.mdcontent as the system prompt - When starting a task that matches a skill, paste the relevant skill content into the conversation
Each AGENT.md contains placeholder conventions. Before using:
- Replace build/test commands with your actual commands
- Add your project's naming conventions and architectural constraints
- Add your team's escalation paths and review policies
- Remove any sections that don't apply to your stack
Steering file vs skill — The AGENT.md is always loaded (standing facts and rules). Skills are loaded on demand (step-by-step procedures). Don't put procedures in the steering file; don't put standing rules in skills.
Every agent needs a human owner — Each AI agent job must have a named person accountable for its outputs. The ai-execution-agent AGENT.md enforces this.
Spec quality is the primary quality lever — Vague specs produce bad agent output at machine speed. The BA and Senior Engineer agents both have spec-writing skills for this reason.
AI-generated code has higher defect rates — 1.7× more major issues, 2.74× more security vulnerabilities than human-written code (CodeRabbit, 2025). The quality-engineer and ai-architect agents are configured with this in mind.
Build loops, not pipelines — Agentic coding works because the development workflow has built-in verification at every step (compiler, tests, type checker). Errors surface immediately and are corrected before cascading. Every agent in this team is configured to run verification after each step, escalate on failure, and close feedback loops back to the spec — not accumulate errors silently to the end.
Contributions are welcome. If you have improvements to existing agents, new skills, or additional research — open a PR.
Good contributions:
- New or improved agent skills based on real workflow experience
- Research reports on agentic development practices (with cited sources)
- Corrections to existing agent rules or skill procedures
- New agent configurations for roles not yet covered
Before contributing, read AGENTS.md — it covers file conventions, commit rules, and how to add agents, skills, and research documents.
This project is MIT licensed. By contributing you agree your contributions are released under the same license.