Multi-Agent Orchestration with ai-devkit + antigravity-kit in Cursor

What is Multi-Agent Orchestration?

Normally you chat with one AI, ask a question, get an answer — done. Orchestration means instead of one AI doing everything, you design a pipeline of multiple agents, each specializing in one thing, where the output of one agent becomes the input of the next.

You give one instruction
         ↓
Orchestrator Agent (the commander)
         ↓
  ┌──────┴──────┐
Agent A      Agent B      Agent C
(plan)       (code)       (test)
  └──────┬──────┘
         ↓
Complete result

Instead of you manually copying results between steps, the orchestrator coordinates everything automatically.

How the Two Kits Map to Orchestration

This is the key insight — the two kits naturally split into two layers:

┌─────────────────────────────────────────────┐
│            ORCHESTRATION LAYER              │
│  ai-devkit — coordinates the overall flow   │
│  /new-requirement → /execute-plan →         │
│  /check-implementation → /code-review       │
└──────────────────┬──────────────────────────┘
                   ↓
┌─────────────────────────────────────────────┐
│              EXECUTION LAYER                │
│  antigravity-kit — specialist execution     │
│  @backend-specialist                        │
│  @frontend-specialist                       │
│  @security-auditor                          │
│  @qa-engineer                               │
│  @debugger                                  │
└─────────────────────────────────────────────┘

ai-devkit = Orchestrator, knows "what to do next" antigravity-kit = Worker agents, knows "how to do it"

Workflow 1 — New Feature End to End

This is the most complete orchestration workflow. You give one instruction at the start, then approve each step as it progresses.

Open Cursor Agent mode, type this prompt:

I want to build this feature: [describe your feature]

Orchestrate in exactly this order:
1. Use /new-requirement to analyze and generate requirements
2. After I approve the requirements, use /plan to create a task breakdown
3. For each task in the plan, use /create to implement it — call the right
   antigravity-kit agent based on the task type (frontend/backend/security...)
4. After the code is done, use /check-implementation to verify against the design doc
5. Use /writing-test to generate tests
6. Finally use /code-review to review everything

Start at step 1. Stop and ask me before moving to the next step.

Why stop at each step? Because good orchestration doesn't mean the AI runs blind. You approve, the AI executes — this is called human-in-the-loop orchestration.

Workflow 2 — Complex Multi-File Debug

Debugging across multiple files needs multiple agents working together: one agent traces the flow, another checks security, another looks at performance.

Orchestration prompt:

Bug: [describe the specific bug]
Files involved: [list the files]

Orchestrate the debug in this order:
1. @debugger: trace the full data flow from entry point to failure,
   list all suspect points
2. For each suspect point, call the right specialist:
   - If it's business logic: @backend-specialist analyzes
   - If it's security-related: @security-auditor reviews
   - If it's UI state: @frontend-specialist investigates
3. Synthesize the root cause and propose a fix
4. After the fix, use /writing-test to write a regression test
   to prevent the bug from coming back

Start at step 1.

Workflow 3 — Auto-Generate Documentation

This workflow uses ai-devkit to orchestrate reading your existing codebase and automatically populating docs/ai/ — extremely useful when applying the toolkit to a project that already has code.

Orchestration prompt:

I just initialized ai-devkit on an existing project. docs/ai/ is empty.
Orchestrate to fill it automatically:

1. Read the full project structure (package.json, folder layout, config files)
2. Fill docs/ai/design/README.md: tech stack, architecture, patterns in use
3. Read the main code files, fill docs/ai/implementation/README.md:
   conventions, naming rules, patterns
4. Read test files (if any), fill docs/ai/testing/README.md
5. Read CI/CD config (if any), fill docs/ai/deployment/README.md
6. After each file, ask me to confirm before writing — I'll correct anything wrong

Start at step 1.

Golden Rules for Orchestration

1. Always use Agent mode — without Agent mode there is no orchestration. The AI cannot read files in Ask mode.

2. One prompt, multiple checkpoints — give one instruction but require the AI to pause and check with you before each critical step.

3. Assign the right task to the right agent:

Task type	Use this
Analyze requirements, planning	ai-devkit `/new-requirement` `/plan`
Implement a specific feature	antigravity `/create` + right specialist
Debugging	antigravity `/debug` + `@debugger`
Review and verify	ai-devkit `/check-implementation` `/code-review`
Testing	both `/writing-test` + `/test`
Documentation	ai-devkit (reads from docs/ai/)

4. The fuller your docs/ai/, the more accurate the orchestration — the orchestrator reads docs/ai/ before dispatching tasks. If docs are empty, the AI has no context to coordinate worker agents effectively.

Start Here — Keep It Simple

Don't try all three workflows at once. Start with the simplest one:

Open Cursor Agent mode → run the Workflow 3 prompt (auto-generate docs)

This workflow requires nothing from you upfront — the AI reads your existing code and fills in docs/ai/ automatically. Once docs/ai/ has content, the other two workflows will be significantly more accurate because the orchestrator has real context to work with.