multi-agent architecture patterns

Multi-Agent Patterns

Split work across agents only when boundaries improve context, ownership, evaluation, or parallelism.

DIFFICULTY
Advanced
EST. TIME
60 min
UPDATED
2026-08-19
On this page
  1. 01Definition
  2. 02Why it matters
  3. 03Step by step
  4. 04Build it
  5. 05Failure modes
  6. 06Production notes
  7. 07Sources

Learning objectives

  • Recognize useful supervisor, specialist, and evaluator patterns
  • Define contracts and shared state between agents
  • Evaluate whether added coordination produces a real advantage

Prerequisites

Agent loops · Tool use · Evaluation basics

Answer first

Canonical definition

A multi-agent system coordinates multiple model-driven workers with separate roles, context, or tools through explicit contracts and shared orchestration.

Operating context

Why it matters

Separate agents can isolate context and permissions or parallelize independent work, but every handoff adds latency, cost, and a new failure boundary.

Role-play alone is not architecture; each agent should have a distinct mandate that can be tested independently.

Method

Step by step

  1. 01

    Prove the split

    Name the context, permission, ownership, specialization, or parallelism benefit that one agent cannot provide cleanly.

  2. 02

    Define contracts

    Specify task request, evidence, output schema, confidence, error states, and who may retry or reject a handoff.

  3. 03

    Centralize control

    Keep budgets, policy, state, and termination in deterministic orchestration rather than distributed negotiation.

  4. 04

    Compare the baseline

    Evaluate the multi-agent design against a single-agent or workflow baseline for quality, cost, latency, and operability.

Hands-on lab

Add an evaluator agent carefully

Separate output generation and rubric-based review while keeping final release under a deterministic gate.

Deliverable

Two role specs, a handoff schema, disagreement policy, shared trace, and baseline comparison.

What breaks

Common failure modes

  • F1Creating many personas that share the same model, context, and permissions
  • F2Allowing agents to recursively delegate without a global budget
  • F3Using an agent to evaluate criteria that deterministic checks could enforce

Beyond the demo

Production notes

  • Trace each agent separately and preserve correlation IDs across the full run.
  • Restrict credentials per role so a specialist cannot silently inherit the supervisor's authority.

Further reading

Sources

  1. [1]Anthropic: Building effective agents

Related Tenten resources

From capstone to production

A working demo is the beginning of the engineering problem.

Tenten helps teams scope one agentic workflow, integrate real data and permissions, define evals and handoffs, and own the path through launch and on-call.