Choose the workflow around the planning decision
Choose orchestration against the task’s planning needs, authority and cost. Fixed workflows, model-routed branches and adaptive agents offer different controls. Evaluate the additional flexibility before approving it.
Trace the interfaces between applications and records before changing a system.
Specify the task before choosing an agentic design. Identify the steps, decisions, tools and permitted effects. The architecture decision then concerns where fixed orchestration is sufficient and where the model needs discretion to select the next step.
The vocabulary is doing real damage here, so it is worth being precise. A workflow is a system where the steps are written in code and a model is called at particular points inside them. An agent is a system where the model chooses what to do next and when to stop. Anthropic’s engineering note on building effective agents draws the line in those terms, and the practical advice that follows from it — start with the composable pattern, graduate only when the task demands it — has held up better than most of what was written about agents in the same period.
Use the simplest orchestration that meets the tested task requirements. Enumerating branches is a useful design question, but not a complete decision rule. Adaptive planning can justify additional cost and control work when it improves outcomes on the relevant cases.
The test
Approve only the flexibility the task needs
Swipe or scroll for the full diagram →
Does adaptive planning improve accepted completion enough to earn its cost?
Owner: Architecture and business task owners. Threshold: Comparable task results, cost, latency, access and permitted effects.
- Expand
- Approve bounded adaptive planning when evaluated benefit and controls meet the conditions.
- Retain narrower scope
- Retain a fixed or model-routed workflow where it meets the task requirements.
- Gather more evidence
- Evaluate unresolved branches and failure paths before granting more discretion.
Representative method for this article, not a measured deployment result.
Does adaptive planning improve accepted completion enough to earn its cost? Compare the evidence with the stated threshold before expanding, retaining a narrower scope or gathering more evidence.
Reviewed 2026-10-02
Compare a fixed workflow, a model-routed workflow and an adaptive agent under equivalent conditions. Measure accepted completion, cost, latency and failures. A demonstration of visible planning does not establish that adaptive planning is necessary, while an unenumerated branch does not prove an agent will handle it successfully.
Compare the operating requirements
| As a workflow | As an agent |
|---|---|
| Set limits on calls, context, retries and execution time | Set limits on planning, tool calls, spend and execution time |
| Test coded paths and the model behaviour within each path | Test allowed paths and sample adaptive trajectories with stated coverage limits |
| A failure points at a step. Someone can go and read it | A failure points at a trajectory. Reproducing it needs the whole trace |
| The audit answer is the code path plus the inputs | The audit answer is the trace, and only if you kept it |
| A new capability means writing a branch | A new capability means a tool, and every tool widens the attack surface |
The comparison identifies additional requirements for an adaptive design. Both workflows and agents can include probabilistic model outputs, and both need controlled retries, evidence and failure handling. The distinction is the discretion granted to the model, rather than an assurance of determinism on one side.
Tools introduce authority and effects that need controls at the receiving boundary. An agent cannot safely be assumed to use every tool in any order: the design can constrain calls, parameters and state. A fixed workflow also needs to validate model outputs before acting. Assess irreversibility alongside access, information exposure and recovery.
If it is genuinely an agent
Record the following operating conditions before accepting the agentic scope. The delivery team must show where the conditions are enforced and how failures reach a responsible owner.
Define a completion condition that can be checked against the task. A schema, validator or system-of-record check may supply part of that evidence. Where completion requires judgement, provide a review route and a bounded stopping rule. A step ceiling limits execution without establishing that the work is complete.
What this does not tell you
Agents may suit work that needs planning in response to intermediate findings. Their value remains a local evaluation question. Open-ended research or cross-system triage still needs bounded tools, access and acceptance conditions rather than an assumption that fixed workflows are always worse.
The choice can be revisited. An instrumented workflow can reveal cases outside its planned branches, while an agent’s observed trajectories can reveal steps worth fixing in code. Use that evidence to change the design instead of treating either pattern as a permanent commitment.
The architecture owner should ask which decisions need model discretion and what improvement that discretion produces. Compare the options against the same task and operating limits. Approve the demonstrated flexibility with its controls, or retain the simpler workflow where it meets the need.