Vice President · TheGreyMatter.ai
In brief
Use a workflow when the permitted path is known. Consider an agent when the next useful step depends on evidence discovered during the task, and keep its authority bounded.
Key takeaways
- An LLM inside a process does not make the process an autonomous agent.
- Separate flexible investigation from consequential execution.
- Compare designs on the same tasks and include review work in the result.
An invoice arrives with a confusing description. A model reads it, software checks the purchase order, and a person approves an exception. Is that an agent? The useful answer depends on who chooses the next step. This guide uses the distinction in Anthropic's architecture guidance: a workflow follows predefined orchestration, while an agent directs more of its own process at runtime. Other teams may use the labels differently.
The distinction is control, not the number of steps
A workflow may use several model calls, route between branches, or repeat a check. Those features do not alone make it an agent. The defining question is whether software has specified the permitted transitions or the model is deciding how to proceed. Anthropic's guide describes both patterns and recommends beginning with simple designs before adding complexity.
OpenAI's practical agent guide likewise distinguishes systems that direct execution from a single model call or classification task. These are design conventions, not a certification standard. In a review, ask to see the actual control flow, available tools, stopping conditions, and escalation path instead of accepting the word agent as a sufficient description.
| Pattern | Who selects the path? | Useful starting point |
|---|---|---|
| Predefined workflow | Application logic within specified branches | Known sequence with explicit exceptions |
| Bounded agent | Model chooses among permitted actions | Investigation whose next question depends on new evidence |
| Hybrid | Workflow controls the stages; an agent investigates a bounded part | Flexible research followed by controlled approval |
Worked example: investigate an invoice discrepancy
In a hypothetical workflow, software extracts the invoice identifier, checks arithmetic, compares the purchase order, and routes a mismatch to an employee. A model may interpret free text, but it cannot invent a new payment action. Missing information follows an explicit exception route. This design is a useful baseline when most cases follow a stable sequence.
A bounded investigation agent could handle the narrower problem of an unclear mismatch. It may search approved purchase orders, delivery records, and correspondence, choosing the next read based on what it finds. Its output is an evidence packet and an explanation of uncertainty. Payment release remains outside that investigation component. The agent's flexibility concerns how it gathers evidence, not whether it may move money.
Write a control contract before giving access
For this example, define the investigation contract in ordinary language before implementation: examine only the selected invoice and permitted related records; stop after a stated time or tool budget; return source references; escalate unresolved contradictions. Make the application enforce the access boundary even if the model requests something outside it.
Decide what happens after interruption. If the task restarts, should it repeat a read, reuse a verified result, or ask for review? Record completed operations separately from proposed ones. A user should not have to infer whether the system merely suggested a change or actually performed it. For any consequential action added later, require a separate, explicit execution policy.
- Allowed inputs and the exact business scope.
- Tools the component may call and the permissions each uses.
- Time, cost, and retry limits that cause a controlled stop.
- Evidence required before returning a conclusion.
- The person or queue responsible for an unresolved case.
Compare both designs on the same cases
Build a small evaluation set containing clean invoices, ambiguous descriptions, missing receipts, duplicate documents, and contradictory correspondence. Keep the expected outcome and acceptable evidence separate from the implementation. Run the fixed workflow and the bounded agent against the same cases and have reviewers use the same acceptance criteria.
Record correct classifications, unsupported conclusions, unresolved cases, elapsed time, tool cost, and human correction effort. Inspect difficult cases individually. An agent that saves time on ordinary invoices but confidently mishandles exceptions may be a poor fit for the task. A workflow that escalates appropriately may create more operating value even if its output looks less autonomous.
Use a decision rule for expansion
Our proposed rollout rule is to expand only when the agent repairs a named class of workflow failures without introducing unacceptable new behavior. Keep the known workflow available as a fallback, and define which exceptions require human handling. An uncertain result should return to that controlled path rather than trigger unlimited additional attempts.
Document why model-directed investigation was selected, what evidence supported the choice, and which change would trigger reevaluation. A new tool, a broader document collection, or permission to write records changes the risk and the test plan. The architecture is a decision about a particular job under particular constraints, not a verdict that every business process should become an agent.
From reading to doing
Put it to work: compare two ways to finish one task
Take a narrow task such as preparing an invoice exception summary. Before building either approach, write what a correct result contains: the invoice identifier, the unresolved discrepancy, the supporting records, and the person who should review it. An attractive explanation without the required evidence is not a completed task. Keep this acceptance rule fixed while comparing approaches.
For a hypothetical trial, include a straightforward mismatch, a missing receipt, and a contradiction between two records. A fixed workflow can gather prescribed fields and route exceptions. An agent can be allowed to choose among a small set of read-only investigation tools. Neither version needs payment authority to test whether investigation improves. The worksheet makes the comparison explicit; it does not supply a benchmark or a universal sample size.
- Write the required output, forbidden actions, and conditions that must trigger a human handoff before running either version.
- Use the same source snapshot and cases for both approaches. Keep expected results available to reviewers, separate from model inputs.
- For each run, record accepted or rejected, supporting evidence, correction minutes, elapsed time, and tool usage. Explain the rejection rather than recording only a score.
- Inspect every case where the two approaches disagree. Identify whether the cause was missing evidence, routing, tool choice, or interpretation.
- Write a limited rollout decision and a stop condition. Preserve a fallback route for any exception class the trial did not adequately test.
Common questions
Can a workflow use a language model?
Yes. A workflow can use a model for extraction, classification, drafting, or routing while application logic still governs the permitted sequence.
Are agents always better for complex work?
No. Complexity alone does not establish that model-directed execution helps. A stable process with explicit rules may work better as a workflow. Compare quality, exceptions, cost, and oversight on the actual task.
Can an agent be read only?
Yes. An investigation agent can choose among permitted searches and return evidence without changing business records. Its actual permissions should be enforced outside the model.
Sources and evidence
AI assisted editorial analysis. Primary sources checked October 10, 2026. The worked examples and proposed review procedures are editorial illustrations, not reported customer results, vendor guarantees, or claims of personal experience.
One useful next step: Choose one idea from this note and test it at the smallest scale that could teach you something this week.
