Key takeaways
- Choose architecture one workflow step at a time. Most useful production workflows mix deterministic code, bounded model judgment, small agent loops and human decisions.
- A fixed path with an exact output belongs in code. A model can interpret one messy input without controlling what happens next.
- An agentic loop earns its complexity only when evidence discovered during the task changes the useful next action and a tested rubric can tell the loop when to stop.
- A subjective pass condition or a decision that requires personal authority is a hard stop. Observe first or keep the named human decision.
Classify one workflow step
Recommendation · diagnostic score 3
Use bounded model judgment inside a fixed workflow
Code should own the route. Give the model one interpretation or draft task with a schema, validator and named fallback.
Minimum implementation contract
- Keep the next step and allowed labels fixed in code.
- Validate schema, sources and forbidden content before continuing.
- Route ambiguous or failed outputs with the original input attached.
Decision trace
- fixed path contributes 0 variance points.
- structured output contributes 1 variance points.
- rubric validation contributes 1 variance points.
- moderate consequence contributes 1 control points.
- The step needs no external tool choice.
How the decision tree works
The diagnostic begins at zero variance. A fixed path, exact output and exact validator add no points. A bounded path adds two; a path discovered during the run adds four. Structured or open outputs add one or two. A rubric adds one, while a subjective pass condition adds three. Moderate and high consequences add one and three control points. Unknown tool choice adds two more.
The number explains how much uncertainty is present, but it does not select the result by itself. Two hard rules override it. If policy requires a named person to make the decision, the tool returns a human gate. If success is still subjective, it returns observation mode because no autonomous system has a dependable stopping or evaluation condition.
| Result | Who chooses what happens next | Minimum pass condition |
|---|---|---|
| Deterministic code | Versioned conditions and state transitions | Exact expected output or state |
| Bounded model judgment | Code owns the route; a model produces one schema-bound interpretation | Schema, source and task-specific validators |
| Bounded agentic loop | A model chooses among allowlisted actions from live evidence | Outcome rubric, budgets, forbidden actions and stop reason |
| Human decision | The person with the required authority | Recorded decision, reason and verified authority |
| Observe first | Nobody executes automatically yet | Agreement on a tested release gate or rubric |
Why the whole workflow should not receive one label
Microsoft's workflow guidance describes a spectrum: code can control the graph while individual executors use model reasoning where it helps. Anthropic draws the architectural boundary in a similar place. A workflow follows predefined code paths; an agent dynamically directs its own process and tool use. Both sources recommend adding complexity only when simpler patterns fall short.
That distinction disappears when a team asks whether an entire onboarding, support or finance process is agentic. One run may contain exact arithmetic, a message classification, a short evidence search and an approval. Giving all four to an agent hides settled rules. Forcing all four into code hides the genuine ambiguity. Use the Workflow Mapper first, then run this decision tree on every step whose performer is disputed.
Four steps that look similar but need different performers
Calculate the response deadline: the plan, timestamp and service rule are known, and code can prove the answer. Deterministic.
Classify the request: the route and allowed labels are fixed, but the message is messy. Bounded model judgment inside the fixed route.
Investigate a delivery failure: the next useful lookup depends on what the previous system returns. A small agent loop may search allowlisted sources, provided it has a success rubric and action budget.
Promise compensation outside policy: a named service owner must decide. Software can assemble the case, but the final commitment remains human.
Turn the result into a testable implementation boundary
- 01Challenge the declared pathWalk recent normal and failed cases. A path is not fixed merely because the procedure omits its exceptions.
- 02Separate release gates from quality scoresForbidden actions, permission limits and required sources are pass or fail. Style and usefulness may be graded without compensating for a broken gate.
- 03Test the simpler result firstCompare deterministic or single-model performance with the agentic proposal on the same cases. Keep added autonomy only when it improves the operating outcome.
- 04Apply action-level risk controlsUse the Automation Risk Assessment on every step that sends, changes, grants, publishes or pays.
- 05Codify repeated choicesWhen the same successful agent choice becomes a stable rule, move it into code and shrink the loop's authority.
Method and source boundaries
The decision rules were checked on 2 August 2026 against Anthropic's distinction between predefined workflows and dynamically directed agents, Microsoft's step-level workflow spectrum and NIST's lifecycle approach to mapping, measuring and managing risk. The tool does not compare vendors, models, prices or benchmark performance.
The point system is an original, inspectable routing aid. It has not been calibrated as a maturity scale. Its useful output is the recommended control boundary and the questions it exposes, not a high or low number.
Limitations and when not to use this
- The recommendation inherits every wrong assumption entered by the user. Verify the current process with real cases and operators.
- The tool does not assess model quality, security, privacy, legal obligations, cost or product compatibility.
- A deterministic design can still encode the wrong rule or fail silently. Test states, exceptions, retries and effects.
- Inputs run in the browser and are not submitted to all-agents. Avoid entering confidential records, personal data or credentials into any planning tool.
Sources
- Building effective agents — Anthropic Accessed 2 August 2026
- Workflows — Microsoft Learn Accessed 2 August 2026
- AI Risk Management Framework Core — NIST Accessed 2 August 2026
Map the whole workflow
Add the classified step to a map with its owner, input, output and exception destination, then repeat for the next step.
Map the whole workflowUli Prantz
Builds and operates all-agents
Uli Prantz builds all-agents, the process-automation platform this site documents. He writes about the operational side of automating recurring business work: where deterministic code beats model judgment, where it does not, and where a human still has to approve.