← Autonomous AI Agents
0/15 complete

Module 03 · Swarm Orchestration

Designing a Manager-Worker Agent Pattern

25 minfocused lesson5practical steps4grounded questions3source links
Open lesson + course map

On this lesson

Course outline

A manager-worker system separates decomposition and reconciliation from bounded task execution. The manager should assign typed work units and apply acceptance rules; it should not become an all-powerful agent with every credential.

After this lesson, you can define task envelopes, worker leases, result schemas, and a controlled commit stage.

// concept

Create the Task Envelope

// json10 lines
{
  "task_id": "catalog-2026-0719-0042",
  "tenant_id": "shop_12",
  "task_type": "review_product_copy",
  "input_refs": ["sku:A14", "policy:v8"],
  "allowed_tools": ["read_product", "read_policy"],
  "deadline": "2026-07-19T12:00:00Z",
  "max_attempts": 2,
  "output_schema": "product_review_v2"
}

Reference authorized data rather than copying a customer database into each task. Sign or protect queue messages where the infrastructure requires it.

// concept

Lease Work Safely

A worker claims a task for a limited lease. It heartbeats while active. If the lease expires, another worker may retry with the same task ID and idempotency keys. Workers cannot create new task types or expand tool scope.

Return status, output, evidence references, usage, and error class:

// prompt — copy me1 line
COMPLETED | NEEDS_REVIEW | RETRYABLE_FAILED | PERMANENT_FAILED

The manager validates schema and evidence. A separate commit service applies approved changes with its own authorization.

// concept

Prevent Recursive Explosion

Set maximum depth, number of workers, tasks per run, turns per worker, tool calls, tokens/cost, elapsed time, and retries. A worker cannot spawn more workers unless explicitly allowed by a bounded plan. Reject duplicate task identities.

// worked_example

Worked Example

A Rawalpindi accounting firm reviews monthly expense descriptions. The manager creates one task per document reference; workers can read a redacted document and account-code policy, then propose a category with evidence. They cannot post to the ledger.

The manager validates every output. High-confidence routine proposals enter a reviewer queue; low-confidence or policy-conflict cases go to an accountant. Only the accounting service posts after approval. If a worker lease expires, another worker resumes the same task without duplicating the ledger entry.

// failure_cases

Failure Cases to Diagnose

7 cases to diagnose

  • Manager holds all secrets

    issue scoped identities at execution time.

  • Work item contains raw sensitive documents

    use authorized references.

  • Two workers commit the same result

    separate proposal from idempotent commit.

  • Worker creates arbitrary subtasks

    enforce allowed types and depth.

  • Manager merges prose inconsistently

    validate a versioned schema.

  • Lease expires during a long tool call

    heartbeat and make effects idempotent.

  • No terminal failure

    cap attempts and route to a human.

// pakistan_angle

Pakistan Angle

For accounting, payroll, medical, legal, lending, hiring, or public-sector work, workers should prepare evidence, not make final professional or rights-affecting decisions. Obtain the appropriate qualified reviewer and preserve the source record.

Queue systems must tolerate intermittent services without double-posting payments, orders, or customer messages. Use PKT deadlines for human operations but UTC timestamps internally, and display both where cross-border staff are involved.

// hands_on

Hands-On Exercise

5 steps

  1. Define a versioned task envelope and result schema.

  2. Specify lease, heartbeat, retry, and terminal failure.

  3. Separate manager, worker, reviewer, and commit permissions.

  4. Add depth, count, cost, and time budgets.

  5. Test lease expiry, duplicate task, invalid output, and partial commit.

// completion_rubric

Completion Rubric

6 checks — tick as you verify

0/6

// sources

Sources

3 official sources — check every claim yourself

// check_yourself

Check yourself

4 questions · answers and options are taken word-for-word from this course

0/4

  1. 1 / 4 · diagnose

    Your work shows this failure mode: “Work item contains raw sensitive documents.” What does the lesson tell you to do about it?