Skip to content

Latest commit

 

History

History
204 lines (151 loc) · 5.51 KB

File metadata and controls

204 lines (151 loc) · 5.51 KB

Lab 06 — Agent Orchestra: Design a Multi-Agent Delivery Workflow

Purpose

Lab 06 brings together everything from Labs 00 through 05. Participants design a complete multi-phase Claude Code workflow that can take a real piece of work from intake through planning, implementation, review, testing, and PR handoff.

This lab brings it all together.

Core Question

If you had to run this work again next week, how would you design the agentic workflow so it is faster, safer, and more reusable?

What Gets Reused

Prior lab What gets reused in Lab 06
Lab 00 local setup, repo commands, toolchain assumptions
Lab 01 primitive decision tree: prompt, tool, skill, agent, hook, MCP, workflow
Lab 02 Implementation Plan, context map, open-question register, work packages
Lab 03 Git history and issue investigation steps
Lab 04 dynamic data sources: logs, tests, CLI output, screenshots, MCP, APIs, docs
Lab 05 diff review, adversarial review, tests, verification evidence

Workflow Output

Participants produce a named multi-agent Workflow with the following sections:

  • workflow name
  • goal
  • use when
  • do not use when
  • inputs
  • data sources
  • skills used
  • tools used
  • MCP servers or external context
  • agents
  • workflow phases
  • handoffs
  • verification gates
  • stop conditions
  • human approval checkpoints
  • output artifacts
  • failure modes
  • reuse opportunities

Output File

Create a single Workflow document that another team member could run without additional explanation.

Example Workflow

Use safe-feature-slice-delivery-example.md as a reference. It demonstrates one acceptable way to combine planning, investigation, dynamic data, implementation, review, testing, and PR handoff into a named workflow.

Required Workflow Phases

Every Lab 06 workflow must include:

  1. intake and goal clarification
  2. context discovery
  3. intent recovery
  4. implementation planning
  5. work disaggregation
  6. implementation
  7. code review
  8. testing and verification
  9. PR handoff
  10. post-run summary

Required Named Agents

Every workflow must define at least:

  • Orchestrator Agent
  • Context Scout
  • Plan Critic
  • Implementation Agent
  • Diff Reviewer
  • Test Strategist
  • Verification Runner

Additional suggested agents:

  • Commit Archaeologist
  • PR Narrator
  • Workflow Librarian

Agent Definition Template

For each agent, define:

  1. mission
  2. inputs
  3. context to gather
  4. tools allowed
  5. tools disallowed
  6. output artifact
  7. stop condition
  8. handoff target

Example Named Agents

Suggested roles:

Agent Role
Orchestrator Agent Owns the workflow, delegates tasks, controls phase transitions, and produces final synthesis
Context Scout Performs bounded codebase Q&A and file discovery
Commit Archaeologist Uses Git history, blame, issues, and PRs to recover intent
Plan Critic Red-teams the implementation plan before execution
Implementation Agent Makes scoped code changes phase by phase
Diff Reviewer Reviews the diff for scope drift and correctness
Test Strategist Designs test strategy from acceptance criteria
Verification Runner Runs checks, captures output, and reports pass/fail evidence
PR Narrator Produces the final PR summary, risks, and follow-ups
Workflow Librarian Identifies what should become a skill, command, hook, or reusable workflow

Handoff Rule

Pass artifacts, not transcripts.

Each workflow should define handoffs with:

  • what is passed
  • what is intentionally not passed
  • the required artifact
  • the success condition for the receiving agent

Example Orchestrator Prompt

Use this as the final execution prompt skeleton:

You are the Orchestrator Agent for the <workflow name> workflow.

Goal:
<goal>

Inputs:
<inputs>

Agents available:
<agents>

Workflow phases:
<phases>

Rules:
- Do not skip gates.
- Use bounded context discovery.
- Pass compact artifacts between agents.
- Stop on blocking questions.
- Implement one phase at a time.
- Run verification before PR handoff.

Begin with Phase 1: Intake.

Exercises

  1. Name the workflow for a real class of work.
  2. Select and define the named agents.
  3. Map the dynamic data sources the workflow needs.
  4. Define artifact-based handoffs between agents.
  5. Define at least three gates: plan review, scope drift, and verification.
  6. Produce the final Orchestrator Agent execution prompt.

Recommended Agent Set for Self-Study

If you do not have a real team workflow to model yet, start with:

  • Orchestrator Agent
  • Context Scout
  • Commit Archaeologist
  • Plan Critic
  • Implementation Agent
  • Diff Reviewer
  • Test Strategist
  • Verification Runner
  • PR Narrator

Pass Criteria

A Lab 06 workflow passes if:

  • it names the workflow
  • it defines at least seven agents
  • each agent has a mission and output artifact
  • it includes dynamic data sources
  • it uses at least one skill
  • it uses at least one tool or command
  • it includes at least three verification or review gates
  • it includes stop conditions
  • it defines handoffs
  • it produces PR-ready output
  • it identifies reuse opportunities

Self-Study Note

If you are taking this course without an instructor, do not skip the artifact collection from Labs 02 through 05. Lab 06 only works when the earlier artifacts are treated as real workflow inputs instead of hypothetical placeholders.

Suggested Reflection

Which agent in your workflow is most likely to become a reusable organizational asset, and what would it need to become stable enough for reuse?