AI agent vs workflow: the difference and when to use each

A workflow runs steps you fixed in advance, in the same order every time; an AI agent gets a goal and a set of tools and decides the steps itself while it runs. Use a workflow when you know the steps and want predictable cost and output. Use an agent when the steps depend on the input and can't be written down ahead of time.

The definitions and trade-offs below come from Anthropic's "Building effective agents", read on 2026-09-28. How Sume offers each option comes from Sume basics, Agent Completions and Errors and spend.

What is the difference between an AI workflow and an AI agent?

Who decides the next step. Anthropic draws the line this way: workflows are "systems where LLMs and tools are orchestrated through predefined code paths", while agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks" (Anthropic).

In a workflow, the path is in your code or your automation tool, and a model fills in one step at a time. In an agent, the path is in the model: it calls a tool, reads the result, and picks the next call until the task is done or a stopping condition ends it.

When should I use a workflow instead of an agent?

Anthropic's advice is to start with the simplest thing that works and add complexity only when needed. Its trade-offs:

• Workflows offer predictability and consistency for well-defined tasks.
• Agents fit open-ended problems where you can't predict the number of steps or hardcode a fixed path.
• Agentic systems often trade latency and cost for better task performance.
• An agent's autonomy means higher costs and the potential for compounding errors, so test it in a sandbox and add guardrails.
• Stopping conditions, such as a maximum number of iterations, keep an agent under control.

What is an agentic workflow?

A fixed workflow with one or more agent steps inside it. Anthropic groups workflows and agents together as "agentic systems", and workflow tools now ship agent steps: n8n's docs describe an AI Agent node you integrate into your workflows (n8n). The workflow still owns the order of steps; the agent step decides how to do its own part.

A saved recipe that an agent follows is another middle ground: the steps are written down, but an agent carries them out and handles what varies from run to run.

How does Sume offer workflows and agents?

At both ends and in the middle. Models are atomic endpoints you chain yourself, which is a workflow. Agent Completions run the Sume agent on a task you send each time. A Format is the middle: a saved recipe the agent follows on every run, where only the inputs change. All three agent surfaces run the same agent and return the same receipt shape. What is a video agent? describes the agent itself.

How do I keep an agent's cost bounded?

Give it a ceiling. Every Sume Format carries a spend cap (generation_spend_cap_usd_micros, $400 when never set), a request can set its own generation_spend_cap_usd up to the $500 platform maximum, and a run that would spend past its cap ends failed. On Agent Completions the cap is required: generation_spend_cap_usd has no default, and a request without it fails with 400 invalid_request. Spend caps for unattended AI agents covers the details. A plain workflow's cost is simply the sum of the calls you wrote into it.