Quick Start¶
Get productive with Agent Flow in 5 minutes.
Installation¶
Or clone and load it directly:
Your First Orchestration¶
Step 1: Start a Session¶
Navigate to your project:
Step 2: Run Your First Task¶
Working in a fresh project with no existing code? Try the first orchestration walkthrough instead — the example above assumes an existing form to modify.
Watch as: 1. Riko explores your codebase 2. Senku creates an implementation plan 3. Loid implements the changes 4. Lawliet reviews code quality 5. Alphonse verifies tests and build
Depending on the profile, Riko and Senku may be skipped; Lawliet and Alphonse always run, in parallel, and Codex joins under standard/thorough if installed. Complex tasks pause once for plan confirmation.
Step 3: Review the Result¶
When you see:
Your task is complete with: - All tests passing - No type or lint errors (checked by Lawliet) - Build successful
If review findings remain after the round cap, the report lists them under "Open findings" and no completion marker is printed.
Deep-Dive for Context¶
For larger tasks, gather context first:
Then use that context:
Quick Reference¶
Commands¶
| Command | Purpose |
|---|---|
/orchestrate <task> |
Execute a complex task |
/orchestrate --use-deep-dive <task> |
Execute with existing context |
/team-orchestrate <task> |
Deprecated — forwards to /orchestrate |
/deep-dive |
Explore entire codebase |
/deep-dive --focus=<path> |
Explore specific area |
/deep-dive --refresh |
Update existing context |
/agent-flow:explain <topic> |
Generate an interactive HTML explainer for a codebase topic |
/agent-flow:analyze |
Load session transcripts and report on subagent behavior |
Agents¶
| Agent | Role | When Active |
|---|---|---|
| Riko | Explorer | Finding files, understanding patterns |
| Senku | Planner | Creating implementation strategy |
| Loid | Executor | Writing and editing code |
| Lawliet | Reviewer | Checking code quality |
| Alphonse | Verifier | Running tests and validation |
| Speedwagon | Author | Writing /explain modules (not part of /orchestrate) |
State Files¶
| File | Purpose |
|---|---|
.claude/orchestration.local.md |
Tracks orchestration progress |
.claude/deep-dive.local.md |
Stores codebase context |
You don't need to edit .gitignore. agent-flow adds a managed block to it the first time a command runs. Set AGENT_FLOW_NO_GITIGNORE to any non-empty value to opt out.
Example Workflow¶
# Start new session
claude --plugin-dir ~/agent-flow
# Gather codebase context (optional but recommended)
/deep-dive
# Implement features using context
/orchestrate --use-deep-dive Add user authentication
/orchestrate --use-deep-dive Add password reset
/orchestrate --use-deep-dive Add email verification
# Refresh context after major changes
/deep-dive --refresh
Tips for Success¶
-
Be specific - "Add JWT authentication with refresh tokens" > "Add auth"
-
Use deep-dive first - For unfamiliar codebases, the upfront investment pays off
-
Trust the verification - When it says VERIFIED, the code passed all checks (a capped run lists open findings instead)
-
Let the system ask questions - If your request is vague, it will clarify
-
Check the summary - Always review what was implemented and verified
Next Steps¶
- Your First Orchestration - Walkthrough for an empty project
- Installation Guide - Detailed setup instructions
- Using Orchestrate - In-depth orchestration guide
- Using Deep-Dive - Context gathering guide
- The "Subagents LIE" Principle - Why verification matters