Skip to main content

Agentic Development

Cornerstone uses an agentic development workflow where a team of specialized AI agents (powered by Claude) collaboratively build software under human direction.

The Orchestrator and Agent Team​

A human orchestrator (the repository owner) directs the work: defining requirements, launching agents, validating output, and approving releases. The 11-member agent team specializes by function -- product ownership, architecture, implementation, testing, security, UX, and documentation.

Each agent has:

  • A specific system prompt and role
  • Persistent memory across sessions
  • Authority within its domain (architects approve schema, security engineer reviews auth code, etc.)
  • A communication protocol (commits with trailers, GitHub comments with agent tags, specification documents)

How It Works​

The typical workflow for a user story:

  1. Human sets direction -- "We need a photo annotation editor." Needs move to GitHub Issues.
  2. Orchestrator launches planning agents -- product-owner writes stories; product-architect designs schema and API.
  3. Human approves the plan -- Reviews acceptance criteria and technical design.
  4. Orchestrator launches implementation -- Agents implement, test, review, and commit in sequence.
  5. Human validates the feature -- Runs manual tests, provides feedback.
  6. Orchestrator promotes to release -- Merges beta to main, updates docs.

The key rule: the orchestrator delegates, never implements. All code, tests, architecture, and docs go through agents.

Flat Delegation Model​

The orchestrator launches all agents directly. The dev-team-lead plays a coordinating role (writes implementation specs, reviews output, commits code) but does not sub-delegate -- it routes specs to implementing agents, not other agents. This flat model keeps handoffs clear:

  • Spec: dev-team-lead writes a spec from the acceptance criteria and architecture
  • Implement: The relevant agent(s) (backend-developer, frontend-developer, etc.) execute the spec
  • Test: qa-integration-tester and e2e-test-engineer write tests in parallel
  • Review: Agent reviewers (product-architect, security-engineer, etc.) approve per domain
  • Fix: If review finds issues, the implementing agent fixes in-session
  • Commit: dev-team-lead stages, commits with all agent trailers, pushes, and creates the PR

One Story Per Cycle​

Each development cycle completes exactly one user story end-to-end before starting the next. This keeps context focused and boundaries clear. The orchestrator breaks epics into stories; stories flow through implementation to merge; completed stories move to Done; the cycle repeats.

The only exception is bundled small fixes (/develop with multiple small issues), which execute in one cycle but each get their own commit.

Learn More​

  • Agent Team -- Detailed roles and models for all 11 agents
  • Skills -- Skill structure and lifecycle (/epic-start, /develop, etc.)
  • Workflow -- PR review gates and release cycle
  • CI & Guardrails -- Automated quality enforcement