Anthropic: Building effective agents
Primary guidance on simple workflows, tool use, orchestration, and adding autonomy only when it creates measured value.
← Back to Compute College
Compute College
Workflows before agents
Use fixed, composable steps before introducing a dynamic agent loop.
Plain-English definition
A workflow is a designed sequence or graph of steps with known transitions. It can still use models and tools, but the application decides what happens next. Start with a workflow when the task path is understandable and add dynamic decisions only where they create measurable value.
Memory trick: Compose known steps; automate uncertainty selectively.
Why it matters
Workflows expose intermediate results, make retries targeted, and keep permission boundaries clear. They are often easier to operate than an agent whose path changes on every request.
- The design makes a hidden tradeoff measurable before it reaches users.
- A clear boundary reduces retries, unsafe actions, or unnecessary capacity.
- The operating choice should be tested against the workload rather than assumed from a demo.
Simple example
A claims pipeline extracts fields, validates numbers, retrieves policy, checks exceptions, and drafts a review packet. Only the exception explanation uses model judgment; the sequence and approvals remain deterministic.
- The example names the input, decision, and expected boundary.
- A deterministic or human checkpoint owns the part a model should not guess.
- The result can be measured with quality, latency, safety, and cost metrics.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Primary reference
This source explains the method or risk boundary. It does not replace testing the actual model, data, provider, and workload conditions in production.
Source discipline: examples on this page are illustrative; measure the real workload before making a release or capacity decision.
Common mistake
Replacing a clear sequence with an agent can hide avoidable complexity. “Let the model decide” may conceal extra calls and failure paths behind a simpler description.
Practical takeaway
What you can do with this
Draw the fixed graph and mark the one step where dynamic choice is necessary. Measure that version before adding autonomy elsewhere.
- Start with a narrow workflow and explicit failure boundary.
- Instrument the path before optimizing it.
- Review the design whenever model, tool, traffic, or policy changes.
Decision check: does each dynamic decision improve an outcome enough to justify its added calls, state, and risk?
Compute College learning path
AI Engineering
Step 32 of 48: Workflows before agents