Tutorial 5: Multi-Agent Orchestration

AgentRole resources declare an AI agent's identity, capabilities, constraints, and handoff behavior. Combined with Pipeline orchestration, multiple agents can collaborate on complex software delivery workflows -- each agent focused on a specific responsibility, passing structured data to the next.

This tutorial builds a three-agent pipeline: an implementation agent writes code, a review agent inspects it, and a deployment agent ships it.

Prerequisites

Completion of Tutorials 01-04
Familiarity with the AgentRole schema and the Agent Orchestration Specification
A working AI-SDLC environment with the @ai-sdlc/reference package installed

Step 1: Define an AgentRole

Create implement-agent.yaml. The Role-Goal-Backstory pattern gives the agent its identity, while tools and constraints define what it can do:

apiVersion: ai-sdlc.io/v1alpha1
kind: AgentRole
metadata:
  name: implement-agent
  namespace: my-team
  labels:
    role: engineer
    stage: implement
spec:
  role: "Senior Software Engineer"
  goal: >-
    Implement well-tested features based on issue specifications,
    producing clean code that passes all quality gates.
  backstory: >-
    A seasoned engineer with expertise in TypeScript and Python.
    Values small, focused changes with thorough test coverage.
    Follows established patterns and avoids unnecessary abstraction.
  tools:
    - code-editor
    - terminal
    - test-runner
    - git-client
    - file-search
  constraints:
    maxFilesPerChange: 20
    requireTests: true
    allowedLanguages:
      - typescript
      - python
      - yaml
    blockedPaths:
      - ".env*"
      - "infrastructure/**"
      - "*.pem"

Key fields:

spec.role -- The agent's title, used in logs and audit trails.
spec.goal -- A clear, measurable objective guiding the agent's behavior.
spec.backstory -- Additional persona context for LLM-based agents.
spec.tools -- The exhaustive list of tools the agent may invoke. The runtime MUST reject calls to unlisted tools.
spec.constraints -- Hard limits enforced by the runtime: file count caps, test requirements, language restrictions, and path blocklists.

Step 2: Add Handoff Declarations

Handoffs define how an agent transfers work to the next agent. Add a handoffs array to the implement agent so it knows how to pass results to the review agent:

  handoffs:
    - target: review-agent
      trigger: "implementation complete and all tests passing"
      contract:
        schema: "./contracts/impl-to-review-v1.json"
        requiredFields:
          - prUrl
          - testResults
          - coverageReport
          - changeSummary

Each handoff entry specifies:

target -- The metadata.name of the downstream AgentRole.
trigger -- A human-readable condition describing when the handoff fires.
contract -- A reference to a JSON Schema plus the fields that MUST be present. The runtime validates the handoff payload against this schema before accepting the transition.

The contract schema (e.g., impl-to-review-v1.json) defines the exact data structure. See the Handoff Contracts section of the spec for versioning rules and validation requirements.

Step 3: Add Skills for Agent Discovery

Skills let other agents and tooling discover what an agent can do. They power the A2A-compatible Agent Card system:

  skills:
    - id: implement-feature
      description: >-
        Implements a feature from an issue specification, including source
        code, unit tests, and integration tests.
      tags:
        - implementation
        - feature
        - testing
      examples:
        - input: "Implement user authentication with JWT tokens"
          output: "Auth module with login/logout endpoints, JWT middleware, and 95% test coverage"
        - input: "Add pagination to the /api/users endpoint"
          output: "Cursor-based pagination with tests covering edge cases"
    - id: fix-bug
      description: >-
        Diagnoses and fixes a reported bug, adding a regression test to
        prevent recurrence.
      tags:
        - bugfix
        - debugging
      examples:
        - input: "Fix: users API returns 500 when email contains a plus sign"
          output: "Input sanitization fix with regression test for special characters"

Each skill declares an id, a description, optional tags for categorization, and optional examples showing representative input/output pairs.

Step 4: Wire Agents into a Pipeline

Now define a Pipeline that orchestrates three agents in sequence. Create delivery-pipeline.yaml:

apiVersion: ai-sdlc.io/v1alpha1
kind: Pipeline
metadata:
  name: feature-delivery
  namespace: my-team
spec:
  triggers:
    - event: issue.assigned
      filter:
        labels:
          - "ai-ready"

  providers:
    issueTracker:
      type: jira
      config:
        projectKey: "ENG"
    sourceControl:
      type: github
      config:
        repo: "myorg/myrepo"

  stages:
    - name: implement
      agent: implement-agent
      qualityGates:
        - unit-test-gate
        - lint-gate

    - name: review
      agent: review-agent
      qualityGates:
        - code-review-gate
        - security-scan-gate

    - name: deploy
      agent: deploy-agent
      qualityGates:
        - integration-test-gate

Each stage references an AgentRole by its metadata.name. The runtime:

Triggers the pipeline when an issue is assigned with the ai-ready label.
Runs the implement stage -- the implement-agent writes code and tests.
Validates the handoff payload, then runs the review stage -- the review-agent inspects the changes.
After review passes, runs the deploy stage -- the deploy-agent handles deployment.

Quality gates at each stage enforce standards before the pipeline advances.

Step 5: Orchestration Patterns

The Pipeline above uses sequential orchestration -- the simplest pattern. The AI-SDLC spec defines five patterns you can apply depending on your workflow needs:

Sequential

Agents execute in stage order. The output of one becomes the input of the next.

implement --> review --> deploy

Best for: standard feature delivery with clear phase boundaries.

Parallel

Independent stages run concurrently. A synchronization barrier waits for all to complete before the pipeline advances.

         +-- code-agent --+
trigger -|                 |-- combine --> review
         +-- test-agent --+

Best for: tasks where code and tests can be written simultaneously, or multiple independent reviews run in parallel.

Router

A complexity scoring step routes tasks to different agent configurations. Simple tasks go fully autonomous; complex tasks get additional oversight.

routing:
  complexityThresholds:
    simple:
      min: 1
      max: 3
      strategy: fully-autonomous
    moderate:
      min: 4
      max: 6
      strategy: ai-with-review
    complex:
      min: 7
      max: 10
      strategy: human-led

Best for: teams that want low-friction automation for simple changes while maintaining human oversight for risky work.

Hierarchical

A lead agent decomposes a complex task and delegates sub-tasks to specialized agents. The lead aggregates results and ensures coherence.

supervisor
  +-- frontend-agent (UI changes)
  +-- backend-agent  (API changes)
  +-- test-agent     (integration tests)

Best for: large features spanning multiple domains or services.

Collaborative

Agents share context and iterate together in a feedback loop. One agent's output feeds back as input to another until a convergence condition is met.

implement <--> review (iterate until approval)

Best for: exploratory tasks where the design emerges through iteration, or review cycles that require multiple rounds of revision.

See Orchestration Patterns in the spec for detailed characteristics and selection guidance.

Step 6: Execute Orchestration Programmatically

Use the SDK to build agents, create plans, and execute orchestrations:

import {
  AgentRoleBuilder,
  sequential,
  parallel,
  hierarchical,
  executeOrchestration,
  validateHandoff,
  createAgentMemory,
} from "@ai-sdlc/reference";

// Build agent roles
const codeAgent = new AgentRoleBuilder(
  "implement-agent",
  "Senior Software Engineer",
  "Implement features with thorough test coverage",
)
  .tools(["code-editor", "terminal", "test-runner", "git-client"])
  .addHandoff({
    target: "review-agent",
    trigger: "implementation complete",
    contract: {
      schema: "./contracts/impl-to-review.json",
      requiredFields: ["prUrl", "testResults", "changeSummary"],
    },
  })
  .build();

const reviewAgent = new AgentRoleBuilder(
  "review-agent",
  "Code Reviewer",
  "Review code for quality and security issues",
)
  .tools(["code-editor", "file-search"])
  .addHandoff({
    target: "deploy-agent",
    trigger: "review approved",
    contract: { requiredFields: ["prUrl", "approvalStatus"] },
  })
  .build();

const deployAgent = new AgentRoleBuilder(
  "deploy-agent",
  "Deployment Engineer",
  "Deploy approved changes to staging and production",
)
  .tools(["terminal", "deploy-cli"])
  .build();

// Create a sequential plan
const plan = sequential([codeAgent, reviewAgent, deployAgent]);

// Set up the agent map
const agents = new Map([
  ["implement-agent", codeAgent],
  ["review-agent", reviewAgent],
  ["deploy-agent", deployAgent],
]);

// Execute with a mock task function
const result = await executeOrchestration(plan, agents, async (agent, input) => {
  console.log(`Executing ${agent.metadata.name} with input:`, input);

  if (agent.metadata.name === "implement-agent") {
    return {
      prUrl: "https://github.com/org/repo/pull/42",
      testResults: { passed: 100, failed: 0, skipped: 2 },
      changeSummary: "Added user authentication",
    };
  }
  if (agent.metadata.name === "review-agent") {
    return { prUrl: (input as any).prUrl, approvalStatus: "approved" };
  }
  return { deployed: true, environment: "staging" };
});

console.log("Success:", result.success);
for (const step of result.stepResults) {
  console.log(`  ${step.agent}: ${step.state}`);
}

Validate Handoffs

Check that handoff payloads meet contract requirements:

const handoffPayload = {
  prUrl: "https://github.com/org/repo/pull/42",
  testResults: { passed: 100, failed: 0 },
  changeSummary: "Added auth",
};

const error = validateHandoff(codeAgent, reviewAgent, handoffPayload);
if (error) {
  console.error(`Handoff failed: ${error.message}`);
} else {
  console.log("Handoff valid");
}

Agent Memory

Agents can use multi-tier memory for context persistence:

const memory = createAgentMemory({
  agentName: "implement-agent",
});

// Working memory (current task context)
memory.working.set("currentIssue", { id: "ISS-42", title: "Add auth" });

// Short-term memory (recent context)
memory.shortTerm.store({ key: "recentPR", value: { url: "..." }, timestamp: new Date().toISOString() });

Validation

Validate each AgentRole resource against the schema:

npx ajv validate \
  -s spec/schemas/agent-role.schema.json \
  -r "spec/schemas/common.schema.json" \
  -d implement-agent.yaml

Validate the Pipeline resource:

npx ajv validate \
  -s spec/schemas/pipeline.schema.json \
  -r "spec/schemas/common.schema.json" \
  -d delivery-pipeline.yaml

Common validation failures:

Missing required fields (role, goal, tools in AgentRole; triggers, providers, stages in Pipeline).
Empty tools array (must contain at least one item).
Handoff contract missing the required schema field.
Skill entries missing id or description.

Summary

In this tutorial you:

Defined an AgentRole with the Role-Goal-Backstory pattern, tools, and constraints.
Added handoff declarations with typed contracts to enable structured inter-agent communication.
Declared skills so agents can be discovered via A2A-compatible Agent Cards.
Wired three agents into a Pipeline with sequential stage execution and quality gates.
Reviewed the five orchestration patterns -- sequential, parallel, router, hierarchical, and collaborative -- and when to apply each.
Used the SDK to execute orchestration programmatically with mock tasks and handoff validation.

Next Steps

Agent Orchestration Specification -- Full reference for agent roles, handoff contracts, state management, discovery, and security.
Pipeline Specification -- Detailed schema for triggers, providers, stages, and routing.
Glossary -- Definitions for Role-Goal-Backstory, handoff contract, Agent Card, A2A, and other key terms.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Tutorial 5: Multi-Agent Orchestration

Prerequisites

Step 1: Define an AgentRole

Step 2: Add Handoff Declarations

Step 3: Add Skills for Agent Discovery

Step 4: Wire Agents into a Pipeline

Step 5: Orchestration Patterns

Sequential

Parallel

Router

Hierarchical

Collaborative

Step 6: Execute Orchestration Programmatically

Validate Handoffs

Agent Memory

Validation

Summary

Next Steps

FilesExpand file tree

05-multi-agent-orchestration.md

Latest commit

History

05-multi-agent-orchestration.md

File metadata and controls

Tutorial 5: Multi-Agent Orchestration

Prerequisites

Step 1: Define an AgentRole

Step 2: Add Handoff Declarations

Step 3: Add Skills for Agent Discovery

Step 4: Wire Agents into a Pipeline

Step 5: Orchestration Patterns

Sequential

Parallel

Router

Hierarchical

Collaborative

Step 6: Execute Orchestration Programmatically

Validate Handoffs

Agent Memory

Validation

Summary

Next Steps