What Is a Deterministic Workflow?
A deterministic workflow is a process that follows a predefined path, so the same input produces the same steps and the same result every time. Nothing in it decides what to do next at run time. Someone decided in advance, and wrote it down.
Workflows and agents
Anthropic’s 2024 post “Building effective agents” draws a useful line. It calls workflows “systems where LLMs and tools are orchestrated through predefined code paths.” It calls agents “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
The difference is who picks the path. In a workflow, the path is fixed ahead of time. In an agent, the model chooses it as it goes.
One caution on terms. Anthropic’s workflows can still contain model calls. A deterministic workflow in the strict sense has none: the model, if there is one, is used to build or repair the path, not to walk it. The line that matters is predefined path versus model-directed path, not AI versus no AI.
Why it matters
- Cost. An agent that re-derives every step pays for reasoning every time. A fixed path pays once, at design time.
- Reviewability. A fixed path can be read, tested and approved before it runs. An agent’s path exists only after the fact.
- Repeatability. Same input, same result. That is a requirement in a lot of institutional settings.
- Failure. When a fixed path breaks, you know which step. When an agent drifts, you read a transcript.
The tradeoff is real. A fixed path only covers cases someone mapped. Anything new needs a person, or a model, to work out a new route first.
How we use it at Gravitas
Gravitas Technologies describes its approach as reasoning once and executing repeatedly: use a reasoning model to work out how a task gets done, record the result as a map, and run the map from then on. Its first product, CampusWay, runs inside a student’s existing university login and works up through four autonomy levels, from reading facts to preparing a draft, and it does not press the final button. The company’s pages say more: Gravitas Technologies.
What this is not
It is not a claim that agents are bad. Agents are the right tool when you can’t map the task in advance. It is a claim that most institutional tasks, such as forms, registration steps and status checks, can be mapped, and that running a map is cheaper and easier to audit than asking a model to rediscover it each time.
Sources
- Anthropic, Building effective agents, December 19, 2024.
- Gravitas Technologies, gravtechnologies.com, checked October 22, 2026.