Key takeaways
- Begin with the least complex architecture that can complete the work and be evaluated.
- Use deterministic code for stable rules and side effects; reserve model judgment for ambiguity that earns its cost.
- Multiple agents help when work is genuinely separable, parallel, or requires isolated context and authority—not because role names sound organized.
- Benchmark architecture variants on the same cases, including handoff loss, retries, latency, cost, and recovery.
01
Separate the workflow from the cast list
A “researcher agent,” “strategist agent,” and “writer agent” can still be one brittle chain of summaries. Draw the information, decisions, tools, state, and side effects first. Then ask which steps are deterministic, which need model judgment, and which can happen independently.
OpenAI’s practical guidance recommends maximizing a single agent before adding multi-agent complexity. That is not a universal ban; it is a reminder that more coordination creates more handoffs to evaluate and maintain.
02
Use architecture as a testable hypothesis
A deterministic pipeline fits stable transformations and policy checks. A single agent can keep context together while choosing among tools. Multiple agents may help when independent investigations can run in parallel, specialist context must stay isolated, or different trust boundaries require separate credentials.
They may hurt when the task depends on one long chain of evidence, summaries discard crucial detail, agents duplicate calls, or nobody owns the final state. An orchestrator does not remove these risks; it becomes another component that can misunderstand the work.
Swipe to compare every column
| Task shape | Start with | Add complexity when |
|---|---|---|
| Stable ordered steps | Deterministic workflow | A step truly needs contextual judgment |
| Ambiguous task, shared context | Single agent with narrow tools | Tool selection or instructions become unmanageable |
| Independent parallel investigations | Coordinator plus bounded workers | Measured latency or quality improves |
| Different trust domains | Separated services or agents | Isolation is enforced at credentials and data boundaries |
03
Keep handoffs lossless enough to audit
Pass source references, structured findings, unresolved questions, confidence, and constraints—not only a polished summary. Give each worker a narrow output contract and prevent it from taking actions outside its assignment. The final decision should be tied back to primary evidence.
Trace each subtask under one operation ID so duplicate work, circular delegation, and conflicting outputs are visible. Set limits on fan-out, depth, turns, tokens, time, and external calls.
04
Compare on the same jobs
Run the simplest viable architecture and the proposed alternative on identical cases. Compare task success, policy violations, missing evidence, tool errors, recovery, latency, and total cost. Inspect trajectories, because two systems with the same final answer may carry very different operational risk.
Choose the design the team can operate and improve. A diagram with more agents may look sophisticated while the simpler system quietly produces better evidence.
Primary sources and further reading
Use the source material to validate details against your own context and current platform configuration.
- OpenAI: A Practical Guide to Building Agents
- Open Agent Specification: Cross-Framework Comparison of AI Agents
- NIST: Tool Use in Agent Systems
This field note follows the XenGrowth editorial policy: primary sources where available, visible limitations, material review dates, and no invented first-hand experience.
Stay with the problem
