Five agent-loop patterns we keep reaching for.
Planner-executor, ReAct, supervisor-worker, hierarchical, pure tool-calling. When each one fits, when it doesn't, and when to mix them.
We rely on five recurring agent-loop patterns: pure tool-calling for deterministic single-step work, ReAct for short reasoning chains, planner-executor for multi-step workflows that benefit from explicit decomposition, supervisor-worker for parallelizable subtasks, and hierarchical for cross-domain orchestration. Each fits a specific problem shape, and most production systems mix two or three.
There's no universal agent loop. The teams that ship reliably pick the simplest pattern that matches the problem, and resist the urge to wrap everything in a planner just because frameworks make it easy.
01 — Pure tool-calling.
The model receives a prompt, picks a tool, returns a result. No reasoning loop. Best for deterministic single-step tasks: structured extraction, function dispatch, classification routing. Easy to reason about, cheap to run, and trivially testable. If your task fits, don't overengineer.
02 — ReAct (Reasoning + Acting).
The model alternates Thought → Action → Observation steps. Good for short tool-using chains where the next step depends on the last result. Failure mode: runaway loops on ambiguous inputs. Always cap iterations.
03 — Planner-Executor.
A planner agent decomposes the task into discrete steps, an executor agent runs them. Best when the workflow is multi-step and benefits from explicit decomposition — e.g., document analysis, structured report generation. Cleaner traces, easier eval, but more latency and cost than ReAct.
04 — Supervisor-Worker.
A supervisor routes subtasks to specialized worker agents (often in parallel). Best for tasks with clear sub-domains: research with separate web/database/code workers, multi-step content with research/writer/editor specialization. Worth the orchestration overhead when subtasks genuinely benefit from focused context.
05 — Hierarchical.
Multiple supervisor-worker trees nested under a top-level orchestrator. We reach for this when domains don't overlap — e.g., a legal-research tree and a financial-modeling tree both feeding into a single deliverable. Powerful, but the failure surface is larger; instrumenting it well is non-negotiable.
Mixing patterns is normal. The supervisor at the top can be planner-executor; its workers can be ReAct or pure tool-calling. The pattern is a tool, not a religion.
How to pick.
Start with the simplest pattern that could plausibly work. If you can solve the task with pure tool-calling, do it. Move to ReAct only when steps depend on previous results. Move to planner-executor only when explicit decomposition genuinely helps trace and eval. Don't reach for multi-agent until single-agent has clearly hit its ceiling.
Guardrails that survive contact with real users.
Have an AI product
that needs to ship?
Tell us where you are — early concept, broken prototype, or scaling something that already works. We'll come back within 24 hours with a take and a quote.