Reference architectures for AI apps, and when each one fits.

Six shapes most production AI systems take, what each is made of and how it fails first. Each one links the lab that builds it.

Read the six patterns

Each pattern names who decides the next step.

Look inside

Each opens to when to use it, its parts, how it fails first and the lab.

  1. codemodelcodeexception

    1Workflow with one model step

    Your code decides.

    Open the details
    When to use it
    The steps are known in advance and only one needs judgement: triage, extraction, drafting for approval.
    Its parts
    A queue or trigger, Deterministic steps, One model call with a schema for its output, An exception path to a human.
    How it fails first
    The model step returns something the next step cannot parse. Validate its output against a schema and route failures to a human.
    The lab
    Agentic AI System Design for Engineers & PMs
  2. modeltooltooltoolstep limit and budget

    2Single agent with tools

    The model decides.

    Open the details
    When to use it
    The path depends on what the agent finds: research, debugging, open-ended support.
    Its parts
    A model in a loop, A small set of well-described tools, A step limit and a budget, Traces for every run.
    How it fails first
    It loops or wanders. Cap steps and spend, and read traces of failed runs every week.
    The lab
    Build a Mission Control for AI Agents
  3. supervisorworkerworkerworker

    3Supervisor with workers

    A supervisor model decides.

    Open the details
    When to use it
    The work splits into parts that need different tools, permissions or can run in parallel.
    Its parts
    A supervisor that plans and routes, Workers with narrow tools, A written contract for what each worker returns.
    How it fails first
    Context is lost at handoffs. Define what each worker returns, and test the handoff, not only the worker.
    The lab
    Build a Multi-Agent System with MCP
  4. indexmodel12answer with citations

    4Retrieval over your documents

    Your code, then the model.

    Open the details
    When to use it
    Answers must come from your own documents and cite them.
    Its parts
    Parsing that keeps structure, Chunking and embeddings, A vector index with filters, Answer generation with citations.
    How it fails first
    The right chunk is never retrieved. Measure recall on real questions before tuning the prompt.
    The lab
    Build a Second Brain with Claude Code
  5. workflow enginemodeltool stepretryapproval

    5Durable agent with human approval

    The model, inside a workflow engine.

    Open the details
    When to use it
    Runs take minutes to days, call slow systems or need sign-off.
    Its parts
    A durable workflow engine, Tool calls as retryable steps, Approval steps that can wait, Idempotent side effects.
    How it fails first
    A retry repeats a side effect, such as sending an email twice. Make every side effect idempotent.
    The lab
    Agentic AI System Design for Engineers & PMs
  6. agentagentagentpolicyMCPMCPMCP

    6Tools behind a gateway

    Policy at the gateway decides.

    Open the details
    When to use it
    Several teams and agents share tools through MCP.
    Its parts
    MCP servers per system, One gateway in front, Access rules per team and agent, An audit log of calls.
    How it fails first
    An agent can call a tool nobody meant it to. Default to deny and log every call.
    The lab
    Secure AI Agents with Guardrails

What each pattern gives you

  1. When to use it.

    The job shape that fits it.

  2. Its parts.

    What you build and run.

  3. How it fails first.

    The break you will see before any other.

  4. The lab that builds it.

    A recording or an upcoming Live Lab.

Build one of the six live on Thursday

Save a free seat