Autonomous Agent Architectures: Plan-and-Execute, Agent Fleets and the Cost Reality
Agents are more than a single loop. ReAct, plan-and-execute, supervisor-worker patterns, heterogeneous model use, and how to keep cost under control across agent fleets.
Building your first agent is easy: give a model tools, put it in a loop, and it works. Making it reliable, cheap and predictable is an architectural problem. In systems making thousands of calls, which model does which job determines both the quality and the entire bill.
Three core patterns
- The ReAct loop: think, call a tool, observe, repeat. Simple and flexible, but because it carries the whole context each step it grows expensive on long tasks and can drift.
- Plan-and-execute: a capable model produces the whole plan first, then cheaper models carry out the steps.
- Supervisor-worker: a coordinator distributes tasks to specialised sub-agents and merges their results.
Why plan-and-execute is cheap
In ReAct every step goes to the strongest model, and the context grows each time. Plan-and-execute separates the two: expensive reasoning happens once, and repetitive work falls to a cheap model. Against using a frontier model for everything, the saving can be very large; reductions approaching ninety percent are reported in industry write-ups.
Supervisor and workers
When a task parallelises naturally this pattern is strong: each sub-agent focuses on a narrow job, carries its own context, and returns only a summary. Context isolation is one of the most valuable properties in agent systems — noise from one subtask does not pollute the others.
The cost is coordination complexity. When sub-agents return contradictory results, who decides must be defined explicitly.
Heterogeneous model use
| Tier | Model | Work |
|---|---|---|
| Planning | Strong, expensive | Strategy, resolving ambiguity |
| Execution | Mid-tier | Standard steps, tool calls |
| High frequency | Small model | Classification, extraction, formatting |
Using one model is simple but expensive. In a tiered approach the real engineering work is deciding which step belongs to which tier — and measuring it.
Stopping the loop
The most common production failure in autonomous systems is the infinite loop: the agent calls the same tool repeatedly, makes no progress, and cost accumulates. Every agent needs three limits: a maximum step count, a maximum token budget, and a progress check. When there is no progress, stopping beats continuing.
Frequently asked questions
Is a multi-agent system always better?
No. When the task is naturally sequential, a single agent is simpler, cheaper and easier to debug. Multi-agent architectures pay off on genuinely parallel work.
How do I observe agent decisions?
Log every step with its input, chosen tool, output and token cost. An agent you cannot trace is an agent you cannot debug.
Where should human approval sit?
Immediately before irreversible actions: transferring money, sending email, deleting data. Read operations can proceed unattended.
Agent architecture is less about model choice than about designing a division of labour. For more, see context engineering and function calling.
