Orchestrating Multi-Agent Systems
Overview
Design and implement multi-agent systems using AI SDK v5 with structured handoffs, intelligent routing, and coordinated workflows across AI providers. This skill covers agent role definition, tool scoping, inter-agent delegation via handoff rules, and workflow orchestration patterns including coordinator-worker and supervisor topologies.
Prerequisites
- Node.js 18+ and TypeScript 5.0+ runtime
- AI SDK v5 (
npm install ai @ai-sdk/openai @ai-sdk/anthropic @ai-sdk/google)
- API keys for target providers set in environment variables (
OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_GENERATIVE_AI_API_KEY)
- Zod for input/output schema validation (
npm install zod)
- Familiarity with agent-based architecture patterns (coordinator, pipeline, broadcast)
Instructions
- Initialize a TypeScript project with
tsconfig.json targeting ES2022 and moduleResolution bundler
- Install AI SDK v5 core and provider packages for each model backend required
- Define agent roles by creating separate modules per agent, each with a system prompt, model binding, and scoped tool set
- Implement tool functions using
ai.tool() with Zod input/output schemas for type-safe execution
- Configure handoff rules using
ai.handoff() to delegate tasks between agents with clear trigger conditions and context passing
- Build routing logic that classifies incoming requests by topic or intent and dispatches to the appropriate specialist agent
- Wire agents into a workflow using sequential, parallel, or conditional orchestration patterns
- Add state management to persist context across multi-step workflows using a shared context object or external store
- Implement circuit breakers and timeout guards to prevent workflow deadlocks
- Test each agent in isolation, then validate end-to-end handoff chains with representative inputs
See ${CLAUDE_SKILL_DIR}/references/implementation.md for the detailed implementation guide.
Output
- TypeScript agent modules with AI SDK v5 provider bindings and system prompts
- Tool definitions with Zod-validated input/output schemas
- Handoff configuration mapping agent-to-agent delegation triggers
- Workflow orchestration files defining sequential, parallel, and conditional execution paths
- Routing classifier that maps user intents to specialist agents
- Integration test suite covering handoff chains and fallback paths
Error Handling
| Error |
Cause |
Solution |
| Provider configuration invalid |
Missing or malformed API key in environment |
Verify process.env.*_API_KEY values; check provider SDK version compatibility |
| Circular handoff detected |
Agent A hands off to B which hands back to A |
Implement handoff depth counter; set maxHandoffDepth and add a fallback terminal agent |
| Task routed to no agent |
Routing classifier returned no match for input |
Add a default catch-all route; improve classifier training data or keyword coverage |
| Tool access violation |
Agent invoked a tool outside its scoped permission set |
Review tools array per agent; ensure tool names match registered definitions exactly |
| Workflow timeout |
Multi-step workflow exceeded deadline without completion |
Set per-step timeouts with AbortController; add workflow-level deadline and partial-result handling |
See ${CLAUDE_SKILL_DIR}/references/errors.md for the full error reference.
Examples
Scenario 1: Customer Support Triage -- A coordinator agent classifies incoming tickets as billing, technical, or general. Billing queries hand off to a specialist agent with access to Stripe tools. Technical queries route to a code-analysis agent with filesystem read tools. Resolution rate target: 85% automated within 3 handoff steps.
Scenario 2: Research Pipeline -- A sequential workflow chains a web-search agent, a summarization agent, and a report-writer agent. Each agent produces structured JSON output consumed by the next. The pipeline processes 50 research queries per batch with a p95 latency under 30 seconds per query.
Scenario 3: Code Review Multi-Agent -- A supervisor agent distributes pull request diffs to specialized reviewers (security, performance, style). Each reviewer returns findings with severity scores. The supervisor aggregates results into a unified review with prioritized action items.
See ${CLAUDE_SKILL_DIR}/references/examples.md for additional examples.
Resources
- AI SDK v5 Documentation -- agent creation, tool definitions, handoffs
- Zod Schema Library -- input/output validation for tools and flows
- Provider integration guides: OpenAI, Anthropic, Google Gemini
- Coordinator-worker and supervisor orchestration pattern references
- OpenTelemetry tracing for multi-agent observability
1---
2name: orchestrating-multi-agent-systems-2
3description: |
4 Execute orchestrate multi-agent systems with handoffs, routing, and workflows across AI providers.
5 Use when building complex AI systems requiring agent collaboration, task delegation, or workflow coordination.
6 Trigger with phrases like "create multi-agent system", "orchestrate agents", or "coordinate agent workflows".
7
8allowed-tools: Read, Write, Edit, Grep, Glob, Bash(npm:*)
9version: 1.0.0
10author: Jeremy Longshore <jeremy@intentsolutions.io>
11license: MIT
12compatible-with: claude-code, codex, openclaw
13tags: [ai, workflow, agent-orchestration]
14
15---
16
17# Orchestrating Multi-Agent Systems
18
19## Overview
20
21Design and implement multi-agent systems using AI SDK v5 with structured handoffs, intelligent routing, and coordinated workflows across AI providers. This skill covers agent role definition, tool scoping, inter-agent delegation via handoff rules, and workflow orchestration patterns including coordinator-worker and supervisor topologies.
22
23## Prerequisites
24
25- Node.js 18+ and TypeScript 5.0+ runtime
26- AI SDK v5 (`npm install ai @ai-sdk/openai @ai-sdk/anthropic @ai-sdk/google`)
27- API keys for target providers set in environment variables (`OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `GOOGLE_GENERATIVE_AI_API_KEY`)
28- Zod for input/output schema validation (`npm install zod`)
29- Familiarity with agent-based architecture patterns (coordinator, pipeline, broadcast)
30
31## Instructions
32
331. Initialize a TypeScript project with `tsconfig.json` targeting ES2022 and moduleResolution `bundler`
342. Install AI SDK v5 core and provider packages for each model backend required
353. Define agent roles by creating separate modules per agent, each with a system prompt, model binding, and scoped tool set
364. Implement tool functions using `ai.tool()` with Zod input/output schemas for type-safe execution
375. Configure handoff rules using `ai.handoff()` to delegate tasks between agents with clear trigger conditions and context passing
386. Build routing logic that classifies incoming requests by topic or intent and dispatches to the appropriate specialist agent
397. Wire agents into a workflow using sequential, parallel, or conditional orchestration patterns
408. Add state management to persist context across multi-step workflows using a shared context object or external store
419. Implement circuit breakers and timeout guards to prevent workflow deadlocks
4210. Test each agent in isolation, then validate end-to-end handoff chains with representative inputs
43
44See `${CLAUDE_SKILL_DIR}/references/implementation.md` for the detailed implementation guide.
45
46## Output
47
48- TypeScript agent modules with AI SDK v5 provider bindings and system prompts
49- Tool definitions with Zod-validated input/output schemas
50- Handoff configuration mapping agent-to-agent delegation triggers
51- Workflow orchestration files defining sequential, parallel, and conditional execution paths
52- Routing classifier that maps user intents to specialist agents
53- Integration test suite covering handoff chains and fallback paths
54
55## Error Handling
56
57| Error | Cause | Solution |
58|-------|-------|----------|
59| Provider configuration invalid | Missing or malformed API key in environment | Verify `process.env.*_API_KEY` values; check provider SDK version compatibility |
60| Circular handoff detected | Agent A hands off to B which hands back to A | Implement handoff depth counter; set `maxHandoffDepth` and add a fallback terminal agent |
61| Task routed to no agent | Routing classifier returned no match for input | Add a default catch-all route; improve classifier training data or keyword coverage |
62| Tool access violation | Agent invoked a tool outside its scoped permission set | Review `tools` array per agent; ensure tool names match registered definitions exactly |
63| Workflow timeout | Multi-step workflow exceeded deadline without completion | Set per-step timeouts with `AbortController`; add workflow-level deadline and partial-result handling |
64
65See `${CLAUDE_SKILL_DIR}/references/errors.md` for the full error reference.
66
67## Examples
68
69**Scenario 1: Customer Support Triage** -- A coordinator agent classifies incoming tickets as billing, technical, or general. Billing queries hand off to a specialist agent with access to Stripe tools. Technical queries route to a code-analysis agent with filesystem read tools. Resolution rate target: 85% automated within 3 handoff steps.
70
71**Scenario 2: Research Pipeline** -- A sequential workflow chains a web-search agent, a summarization agent, and a report-writer agent. Each agent produces structured JSON output consumed by the next. The pipeline processes 50 research queries per batch with a p95 latency under 30 seconds per query.
72
73**Scenario 3: Code Review Multi-Agent** -- A supervisor agent distributes pull request diffs to specialized reviewers (security, performance, style). Each reviewer returns findings with severity scores. The supervisor aggregates results into a unified review with prioritized action items.
74
75See `${CLAUDE_SKILL_DIR}/references/examples.md` for additional examples.
76
77## Resources
78
79- [AI SDK v5 Documentation](https://sdk.vercel.ai/docs) -- agent creation, tool definitions, handoffs
80- [Zod Schema Library](https://zod.dev) -- input/output validation for tools and flows
81- Provider integration guides: OpenAI, Anthropic, Google Gemini
82- Coordinator-worker and supervisor orchestration pattern references
83- OpenTelemetry tracing for multi-agent observability
jeremylongshore