Two systems can produce the same customer reply yet be built in opposite ways. In one, a developer wrote every step in advance and the model fills in the parts that need language. In the other, the model looks at the situation and chooses what to do next.
The first is an AI workflow; the second, an AI agent. The more flexible option is not automatically the better one. For how an agent's loop works, start with our explainer on what AI agents are.
The difference: who controls the path
Anthropic's engineering guidance draws the line clearly:
“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”
— Anthropic, Building effective agents.[1]
Agents, by contrast, are systems where models “dynamically direct their own processes and tool usage.”[1] LangChain's documentation makes the same split: workflows have “predetermined code paths.”[2]
OpenAI likewise excludes applications that don't let the model “control workflow execution.”[3]
Neither design gives the model unlimited freedom. A workflow can still use a model to classify or grade; an agent works only with the tools it has been given. The question is narrower: when one step finishes, does your code pick the next one, or does the model?
Workflow patterns you can use before reaching for an agent
Anthropic describes five workflow patterns, building up from a single model call enhanced with retrieval, tools or memory.[1] Google Cloud documents close equivalents and says its sequential, parallel and loop patterns run on predefined logic, without a model orchestrating them.[4]
- Prompt chaining. Each call works on the previous call's output, with optional checks in between. Suits tasks that split cleanly into fixed subtasks.
- Routing. A classifier sends each input to a specialised path, such as refunds or technical support.
- Parallelisation. Independent subtasks run at once, or one task runs several times for comparison.
- Orchestrator-workers. A central model decides the subtasks at runtime, then combines the results. Here model judgement starts shaping the path.
- Evaluator-optimizer. One call generates, another critiques, and the loop repeats until clear criteria are met.
Many tasks never need to go further. Anthropic notes that optimising a single model call with retrieval and in-context examples “is usually enough” for many applications.[1] Our explainer on retrieval-augmented generation covers that starting point.
AI Chronicle
A decision guide: four questions
The vendor guidance converges on four practical questions.
1. How predictable must the process be?
Anthropic says workflows “offer predictability and consistency for well-defined tasks,” while agents suit open-ended problems where “you can't hardcode a fixed path.”[1] If the same input must always trigger the same steps, that points to a workflow. Anthropic's evaluation guidance also notes that “model outputs vary between runs,” so it runs multiple trials of each test.[5] Variation is easier to accept when the end result is easy to check.
2. What can each task cost, and how long can it take?
More autonomy usually means more model calls. Anthropic warns that agentic systems “often trade latency and cost for better task performance.”[1] Google Cloud says its predefined sequential pattern “can reduce latency and operational costs” compared with model-orchestrated designs, and adds that “this efficiency comes at the cost of flexibility.”[4]
A workflow's cost per request is easier to forecast. But judge cost per acceptable result: an agent that resolves an awkward case may beat a workflow that sends it to a person.
3. How should errors be caught?
In a workflow, you know where each step's output arrives, so you can put a check there; Anthropic calls these “gates.”[1] An agent's value lies partly in recovering from unexpected results, and that cuts both ways:
“The autonomous nature of agents means higher costs, and the potential for compounding errors.”
— Anthropic, Building effective agents.[1]
The usual controls are stopping conditions, such as a turn limit, and handing control to a person after repeated failures.[1][3] Our guide to why AI agents fail walks through how to locate the layer where an error started.
4. How will you explain what happened?
A workflow's design is its first audit record: the code shows which steps can run, in what order. With an agent, the explanation lives mainly in the transcript, the record of a run's outputs, tool calls and intermediate results.[5] Reviewing those is possible, but you have to plan for it.
Permissions apply whichever design you choose. OWASP advises enforcing authorisation in downstream systems “rather than relying on an LLM to decide if an action is allowed,”[6] and OpenAI recommends human oversight for actions that are “sensitive, irreversible, or have high stakes.”[3] Who answers for a mistake is the subject of our analysis of accountability when an AI agent acts.
| Question | Lean towards a workflow when… | Lean towards an agent when… |
|---|---|---|
| Path | The steps are known in advance | The steps depend on what the system finds |
| Predictability | The same input must follow the same steps | Variation is acceptable because outcomes are checked |
| Cost and latency | Volume is high and the per-request budget is tight | The task's value justifies extra calls |
| Errors | Failures are known and can be caught at fixed checks | Recovery from unexpected results is the point |
| Audit | You must show which step ran by design | You can record and review full run transcripts |
AI Chronicle synthesis of the sources above.
A worked example: one inbox, two designs
This is a hypothetical design exercise, not a tested deployment.
Imagine a small online shop whose customer emails mostly ask where an order is, request a return label or report damage.
The workflow version. A routing step classifies each email into those three categories or “other.” For order status, code looks up the order and fills a reply template; the model only adjusts tone. Returns and damage follow their own fixed steps. “Other” goes to a person, and every reply waits for approval.
This is cheap per email, testable category by category and easy to explain: you can see which branch an email took.
Where it strains. Suppose the “other” pile grows, and those emails need investigation: carrier tracking, warehouse notes, then the returns policy applied to an unusual case. No fixed sequence covers them, because each check depends on the last.
The graduated version. Keep the workflow for the known categories. For “other,” add an agent with read-only tools for tracking, notes and policy, a turn limit, and an instruction to return an evidenced draft or escalate. Refunds and outbound messages stay outside its permissions.
The agent handles only the part with no fixed path. Before switching it on, the shop would test it on past “other” emails against what staff actually sent.
When to graduate from workflow to agent
Anthropic's summary advice is to add “multi-step agentic systems only when simpler solutions fall short.”[1] OpenAI suggests agents for work involving nuanced judgement, rule sets that have become hard to maintain, or heavy reliance on unstructured data, adding: “Otherwise, a deterministic solution may suffice.”[3]
Signals that a workflow has reached its limit include:
- a growing fallback branch that needs case-by-case investigation;
- branching so dense that every new case means editing the flow;
- inputs too varied to route reliably.
When you move, keep tested workflow branches, give the agent the narrowest task and tools that cover the gap, and measure it against the workflow on the same cases.
Questions readers commonly ask
Is an AI workflow just traditional automation?
Not quite. It keeps a fixed sequence but uses a model inside steps that need language, such as classifying or drafting.
Can one system be both?
Yes. Anthropic describes its patterns as building blocks to “shape and combine.”[1] A common design is a workflow that hands one branch to an agent, as in the worked example.
Are agents always more accurate?
No source cited here claims that. Anthropic recommends adding complexity only when it demonstrably improves outcomes.[1] Test on your own cases.
Does a no-code builder decide which one I get?
No. Visual tools can build fixed sequences or nodes where the model chooses among tools. The design question stays the same.
The next step
Pick one recurring task and write down its steps. If you can list them all in advance, build a workflow first. If one part keeps needing judgement about what to check next, that part is your agent candidate. Our tutorial on how to build your first AI agent without code shows that narrow, reviewable first build.
Related: What AI agent demos leave out · Build an ecommerce product-question agent
About these sources
All pages accessed 16 October 2026. Vendor guidance describes how these companies recommend designing systems; it is not independent evidence of performance. The worked example and decision table are AI Chronicle's own illustrative material.
Sources
- Anthropic, Building effective agents
(published 19 December 2024; the page now notes that much of the tooling landscape has changed since). Workflow and agent definitions; the five workflow patterns; latency and cost trade-off; “usually enough”; predictability; compounding errors; stopping conditions; “only when simpler solutions fall short.”
- LangChain, Workflows and agents
LangGraph documentation. Predetermined code paths versus agents that define their own processes.
- OpenAI, A practical guide to building agents
(PDF). Agent definition and exclusion of apps that don't let the model control workflow execution (p. 4); when to build an agent and “a deterministic solution may suffice” (pp. 5–6); exit conditions (p. 14); human intervention after failure thresholds and for high-risk actions (p. 31).
- Google Cloud, Choose a design pattern for your agentic AI system
Cloud Architecture Center (last reviewed 28 May 2026). Predefined sequential, parallel and loop patterns; latency and cost versus flexibility.
- Anthropic, Demystifying evals for AI agents
(published 9 January 2026). Variation between runs and multiple trials; definition of a transcript.
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency
Authorisation enforced in downstream systems; human approval for high-impact actions.
This article was researched and drafted with the help of AI tools, then edited and fact-checked by the AI Chronicle editorial team. Illustrations are original artwork created for AI Chronicle with AI image generation.