Agentic Architecture Patterns
The five composable patterns between 'one prompt' and 'full agent' — and how real products stack them.
Five patterns cover most systems. Prompt chaining: fixed sequence, each step's output feeds the next (outline → draft → edit) — use when the decomposition is known. Routing: classify the input, dispatch to a specialized path — cheap model as receptionist, right prompt for each job. Parallelization: fan independent subtasks out simultaneously and aggregate — sectioning (different parts) or voting (same task, multiple samples, majority/judge picks). These three are workflows: you own the control flow.
Two patterns hand the model more control. Orchestrator–workers: a lead model decomposes the task dynamically and delegates to workers (often cheaper models or parallel instances), then synthesizes — the decomposition itself is model-decided, which is what separates it from static chaining. Evaluator–optimizer: a generator produces, a critic scores against explicit criteria, loop until pass or budget — the workhorse for quality-sensitive output. Cap the iterations; unbounded self-critique loops burn budget chasing marginal gains.
Real systems stack patterns: route first; simple intents hit a single prompt, complex ones hit an orchestrator whose workers use evaluator loops. Multi-agent (multiple persistent agents with roles and handoffs) is the far end of the spectrum — powerful for genuinely parallel research-shaped work, but coordination overhead is real and shared-context bugs are subtle. The design instinct to build: start from the least-agency end and add autonomy only where evals show static control flow failing.