Claude Agent Flow

CLAUDE AGENT FLOW

Plan. Build. Review.

Agentic Workflow System within Claude Code - multi-agent pipeline that takes ideas from plan to production.

Model-matched agentsDeep reasoning where it counts, fast execution everywhere else
Enforced code qualityAdversarial critic loop catches issues before anything ships
Unattended executionHand it a plan or task ticket and walk away
Single-shot deliveryWorking, reviewed, tested, and documented code in one run

12 specialized AI agents. One command. Claude Agent Flow orchestrates planning, building, reviewing, and pushing to production code using Claude Code.

agent-flow — build session
ORCHESTRATORReceived brief: /build @plans/demo.md
EXPLORERScanning codebase — mapping docs/, .claude/agents/, patterns…
ARCHITECTDesign complete. Single-file with embedded JS — no deps.
BUILDERWriting visualiser.html — console, flow diagram, controls…
CRITICReviewing… PASS — engine correct, layout verified.

Pipelines

Agent Roster

Orchestrator
Opus

Master coordinator. Breaks work into subtasks and delegates to specialists; never writes code itself.

Ideator
Opus

Lateral thinker. Explores solution space and stress-tests approaches; output goes to the human, never directly to builders.

🧠 Brainstorming Skill
Researcher
Sonnet

Research specialist. Finds current documentation, compares libraries, and surfaces best practices online.

⌕ WebSearch
Explorer
Haiku

Codebase navigator. Fast, cheap mapping of files, dependencies, and patterns before any task begins.

Architect
Opus

Design authority. Produces written design decisions and file-level plans before any implementation begins.

⌕ WebSearch
Storage
Sonnet

Data persistence specialist. Sole owner of RLS policies; covers Postgres, Supabase, S3, and local filesystem.

Frontend
Sonnet

UI specialist. Owns React, TypeScript, Tailwind, browser APIs, accessibility, and client-side state.

🎨 Frontend Design Skill🌐 Playwright
Backend
Sonnet

Server specialist. Owns API routes, business logic, database queries, auth, and all server-side concerns.

Critic
Opus

Adversarial reviewer. Actively tries to break code with edge cases; returns FAIL with specifics or PASS when solid.

🌐 Playwright
Tester
Sonnet

Test specialist. Writes and runs tests; never marks a task done if tests are red.

🌐 Playwright
Reviewer
Sonnet

Standards checker. Read-only code review returning structured BLOCKER / WARNING / SUGGESTION findings.

🌐 Playwright
Author
Haiku

Documentation specialist. Updates README, docstrings, CHANGELOG after reviewer sign-off.

Opus · Deep reasoning
Sonnet · Balanced
Haiku · Fast & efficient

What's inside — Three building blocks

⌨

Commands

Slash commands like /build, /plan, /review orchestrate the full pipeline — type one word to trigger a complete workflow

🤖

Agents

Specialised AI agents, each with a focused role: Architect designs, Builder builds, Critic challenges, Reviewer approves

✨

Skills

Reusable skill modules injected into agent briefs — design systems, testing frameworks, domain knowledge

Project structure

agent-team/
├── .github/
│   └── workflows/
│       ├── agent-flow-backlog-notify.yml   ← backlog notifications
│       └── agent-flow-review-pr.yml        ← AI PR review
├── .claude/
│   ├── agents/          ← 12 agents
│   ├── commands/        ← /build, /plan, /review…
│   └── skills/          ← skill modules
└── .claude-agent-flow/
    ├── hooks/           ← git & session hooks
    ├── scripts/         ← install & sync scripts
    └── tests/           ← BATS test suite

What's bundled

Tool Purpose Why it's there
Backlog.md CLI Task & backlog management Native task tracking integrated into pipeline
BATS Shell test framework Validates pipeline structure and file integrity
Playwright CLI Browser automation Visual regression and page verification
Mergiraf Merge driver Semantic merge conflict resolution
ShellCheck Shell script linter Keeps hook scripts clean
rsync File sync Plugin and skill sync operations
GNU parallel Parallel execution Speeds up test runs
⚡
Early version — under active development. Agent behaviour, commands, and pipeline steps may change as the system evolves. The site reflects the current state of the main branch.