Patterns: six ways to wire a loop

Six patterns cover almost everything built in production today. They are not frameworks — you can implement any of them in a hundred lines with a model client and a while loop. Choosing well is mostly about two questions: does the next step depend on the last result? and what is the cost of a wrong action?

6.1ReAct, in code

Reasoning and acting interleaved. It is the default because it handles surprise — which is what most real tasks are made of.

trajectory = [system_prompt, goal]

for step in range(MAX_STEPS):
    thought = model(context=trajectory, tools=TOOLS)

    if thought.is_final:
        return verify(thought.answer)          # never trust the claim

    action = parse(thought.tool_call)          # schema-validated
    if not permitted(action):
        return ask_human(action)

    result = run(action, timeout=30)           # errors are data
    trajectory += [thought, result]

    if budget_exceeded() or repeated_failure():
        return escalate(trajectory)

The interesting parts are the lines that are not the model call: schema validation, the permission check, the timeout, the repetition detector, the budget. Those five lines are the difference between a demo and something you'd let touch a customer.

6.2Plan first, when the plan is the product

If a human needs to approve the approach, or you need an audit trail of intent, generate the plan as a separate artifact before doing anything. It costs one extra call and buys you reviewability, parallel execution, and a natural place to attach estimates.

plan = model.plan(goal)          # steps, tools, dependencies, risks, cost estimate
if plan.requires_approval:
    plan = await human.review(plan)   # approve / edit / reject

for step in topo_sort(plan.steps):     # independent steps can run in parallel
    result = run(step.tool, step.args)
    if step.failed:
        plan = model.replan(goal, plan, results, error)  # bounded: 2 attempts

Note the bound on replanning. Unbounded replanning is how an agent spends an afternoon and $40 rediscovering that the API key is expired.

6.3Routing: the cheapest win in the system

Most requests are easy. Sending every one of them to the largest model is the single most common way to waste budget. A router classifies first — by rules, by embedding similarity, or by a small fast model — then hands off.

6.4Choosing: answer two questions