THE SHORT ANSWER
An agent can fit variable, multi-step work that requires choosing among tools as evidence changes. Avoid an agent when rules are stable, exact predictability is required, reliable tools or data are missing, latency and cost must be minimal, or a mistake cannot be contained. Start with the least autonomous design that meets the outcome.
Look for justified variability
| Question | Agent may help | Simpler design may be better |
|---|---|---|
| Path | Next steps depend on new observations | Steps and rules are known in advance |
| Judgment | Inputs are messy and require interpretation | Outputs must follow exact calculations |
| Tools | Reliable read or reversible tools exist | Critical systems lack safe interfaces |
| Verification | Success can be checked from environment state | Correctness is hard to observe |
| Consequence | Errors can be contained or approved | One wrong action creates unacceptable harm |
| Economics | Variability justifies added calls and latency | High volume rewards a fixed efficient path |
Evidence & context: Anthropic
Test simpler alternatives first
- A deterministic rule for exact policy or calculation.
- One model call for classification, extraction or drafting.
- Retrieval plus an answer when the need is information, not action.
- A fixed AI-enabled workflow when stages are known.
- A human-owned process when the decision requires accountable judgment.
This sequence is not anti-agent. It keeps complexity proportional to the unresolved variability. Anthropic's published distinction similarly recommends workflows for predefined paths and agents where model-directed choice is genuinely needed.
Evidence & context: Anthropic
Assess action, not only answer quality
A research agent that proposes sources can be corrected before publication. An agent that sends money, changes access or communicates externally creates a different risk even if its language quality is identical. Consider reversibility, blast radius, sensitive data and who can detect an error in time.
Require human approval when consequences exceed the system's tested ability to decide. Confidence scores can help route cases, but they are not proof of correctness or permission.
Evidence & context: NIST
Use a staged decision
- Define the goal and accepted final state.
- Map fixed and variable decisions.
- Confirm tools, permissions and observable outcomes.
- Estimate full cost, including review and failure.
- Run representative cases in a contained environment.
- Compare against the simplest credible alternative.
- Proceed only if added autonomy improves the measured task.
Sources & further reading
- Building effective agents
Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
- Evaluate a model router
Microsoft Learn. Official guidance for evaluating routing across representative workloads using quality, cost, latency and policy criteria. It is not evidence that routing always improves results.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗