Agent architectures, compared
Single-loop, planner-executor, supervisor and graph — what each shape is good at and where each one falls over.
Agent frameworks multiply faster than the ideas behind them. Underneath, there are a handful of shapes, and picking the right one matters more than picking the library.
Single loop with tools
One model, a set of tools, iterating until done. Simple to build, easy to debug, and adequate for a surprising number of real tasks. Start here. It degrades when the task needs more steps than the model can hold coherently.
Planner and executor
One pass produces a plan, a second executes each step. The plan is inspectable and approvable before anything happens, which matters enormously when actions have consequences. The weakness is rigidity — a plan made before execution cannot react to what execution reveals, unless you allow replanning.
Supervisor with specialists
A coordinator routes work to narrow agents. Genuinely useful when the sub-tasks need different tools or different models. Frequently over-applied: if your specialists share the same tools and prompt, you have added latency and failure modes to buy nothing.
Explicit graph
States and transitions defined in code, with the model deciding within nodes rather than deciding the flow. Less autonomous and far more predictable. For regulated or high-value workflows this is usually the right answer — you keep the reasoning and give up the improvisation.
Picking a shape
The question is how much unpredictability the task can tolerate.
- Bounded task, low stakesSingle loop with tools
- Multi-step with real consequencesPlanner + executor, approval gate
- Genuinely different sub-tasks and toolsSupervisor + specialists
- Regulated, audited or high-valueExplicit state graph
Choosing
The question is how much unpredictability the task can tolerate:
- Bounded task, low stakes → single loop
- Multi-step with consequences → planner-executor with an approval gate
- Genuinely heterogeneous sub-tasks → supervisor
- Regulated, audited or high-value → explicit graph
The part that actually decides success
Good tools
- Narrow and specific, not one flexible query tool
- Descriptions that say when not to use them
- Errors that explain how to retry
- Idempotent where there are side effects
Bad tools
- Two tools with near-identical descriptions
- "Error: invalid input"
- One endpoint that can do anything
- Permissions assumed rather than enforced
The part that decides success
Not the architecture — the tools. Well-designed, narrowly scoped, clearly documented tools with good error messages make a simple agent work. Vague, overlapping tools make a sophisticated one flail. Spend your effort there.