Workflow design
Draft for human review
August 2026 · 8 min read
The most useful first agent may be a workflow that does not look autonomous.
The goal is not to make a system feel clever. It is to make a recurring piece of work clearer, more reliable, and easier for a human team to supervise.
“Add an agent” has become a popular answer before the team has finished asking the question. A more useful first question is: what part of this work actually needs model-led judgment?
Anthropic draws a distinction that helps: a workflow follows predefined code paths for LLMs and tools, while an agent dynamically directs its own process and tool use.[1] Neither is inherently better. The right choice depends on how much of the route is already known—and whether the additional flexibility is worth the added cost, latency, and oversight burden.
01. The distinction changes the build
A workflow can classify an incoming request, retrieve approved material, draft a response, run a programmatic check, and hand a result to a person. It may contain several model calls, but the route between them is designed in advance. That can be a strength when the task is repeatable and the process is legible.
An agent is useful when the route cannot be fully specified in advance: it may need to choose among tools, investigate a codebase, react to intermediate results, or recover from a failed attempt. Anthropic recommends starting with the simplest solution and only increasing complexity when performance requires it; its guide notes that workflows trade more predictable paths against the flexibility of model-led decisions.[1]
Autonomy is not a feature to maximize. It is a cost you pay when a fixed route cannot do the job.
02. Run the task through a fit test
Use these questions before introducing a model that can choose its own next step. They are a practical synthesis for first-pilot planning—not a claim that a yes/no quiz can choose architecture for you.
Is the work repeated?
If the task appears once a year, you may be automating a memory problem rather than a workflow. Start with a better template or a manual runbook.
Can you name the outcome?
If the team cannot say what a correct result looks like, an autonomous loop will create more ambiguity, not less.
Is the route mostly known?
If yes, begin with an explicit workflow. Keep the places where judgment is required narrow and visible.
Does it rely on messy context or exceptions?
This is a sign that a model may add value, but it is also a reason to scope the environment and add review.
Can you observe a bad result before it matters?
If not, the first move may be to make the work reversible or introduce a review boundary—not to give the system more reach.
OpenAI’s practical agent guide says teams should prioritize workflows that have resisted traditional automation, especially where nuanced decisions, exceptions, unstructured data, or conversational interaction are central.[2] That is not a mandate to build an agent. It is a clue that a deterministic process may struggle and a carefully bounded model step may be worth testing.
03. Climb complexity only when evidence demands it
Start at the least autonomous level that could create useful leverage. Then test it against real cases. Only promote the system to the next rung when the current design fails for a reason you can name.
A practical complexity ladder
- 01A clear template or manual runbook.
- 02A single model call with carefully defined input and output.
- 03A fixed workflow with checks between steps.
- 04A workflow with a bounded model decision or routing step.
- 05A tool-using agent in a narrow, reviewable environment.
- 06Broader autonomy only after you have evaluation cases, boundaries, and a reason to need it.
This is not an anti-agent argument. It is an argument for preserving optionality. A structured workflow teaches you what the work contains, gives you traceable failure cases, and can later reveal exactly where an agent would help. OpenAI’s guidance on evals recommends identifying where nondeterminism enters the system and evaluating there.[3] A simpler architecture gives you fewer unknowns at once.
Related decision
Build a five-case evaluation suiteSources and review note
Read the material behind the note.
Fact-check status: source links reviewed against retrieved first-party pages on 22 Aug 2026. This is The Model Note editorial guidance, not architecture, security, or operational assurance. Human editorial approval is required before public indexing.