What is an agentic workflow?
An agentic workflow is a process in which language model calls and tools run along a path defined in code, with deterministic steps handling routing and checks and model steps handling judgement.
The phrase is used loosely. Some people mean any multi-step process that involves a language model. In the labs' own vocabulary the line is sharper: in a workflow, code decides the order of steps; in an agent, the model decides its own next step and which tool to use.
Most production systems sit between the two. A fixed path handles the predictable parts, such as fetching a record, validating a field or routing a request, and a model step handles the parts that need reading or judgement.
Common workflow patterns
Prompt chaining
One model call's output feeds the next, with code checks between steps.
Routing
A classifier step sends each input to the branch built for it.
Parallelisation
Independent subtasks run at the same time, or several attempts run and vote.
Orchestrator and workers
A model step breaks a task into parts and hands them to worker steps.
Evaluator and optimiser
One step drafts, another grades, and the loop repeats until the grade passes.
Workflow or agent
Anthropic, OpenAI and Google all advise starting with the simplest structure that works. When the steps are predictable, a workflow or even a single model call is cheaper, faster and easier to test. An agent earns its place when the number of steps cannot be known in advance, inputs are unstructured, or the task needs judgement a rules engine cannot hold.
Google's post on why it built ADK 2.0 (July 2026) made this explicit with a graph runtime that mixes deterministic routing nodes with model nodes. In Google's refund example the change cut tokens by about 50% and latency by about 20%, and keeping control flow out of the model also limits what a prompt injection can change.
How Motiontic's engineers apply it
Motiontic's engineers map the process first and mark which steps are rules and which need reading or judgement. Rules become code. Judgement becomes a model step with its own eval set. Only the parts that genuinely need open-ended planning are given to an agent loop.
Approval gates, retries and logging sit in the code path, not in the prompt, so they hold even when the model behaves unexpectedly.
An agentic workflow compared with an autonomous agent
| Agentic workflow | Autonomous agent | |
|---|---|---|
| Who decides the next step | Code | The model |
| Fits | Predictable, repeatable processes | Open-ended tasks with an unknown number of steps |
| Cost and latency | Lower and steady | Higher and more variable |
| Testing | Each step can be tested on its own | Graded on outcomes across many runs |
| Failure mode | Breaks on inputs the path did not expect | Wanders, skips steps or stops early |
Common questions
What is the difference between an agentic workflow and an AI agent?
In an agentic workflow, code fixes the order of steps and the model fills in the parts that need judgement. In an AI agent, the model chooses its own steps and tools in a loop until the task is done.
Are agentic workflows cheaper than agents?
Usually, because fewer model calls are spent on routing and error handling. Google reported about 50% fewer tokens and about 20% lower latency when its refund example moved routing out of the model in ADK 2.0.
When should a team use an agent instead of a workflow?
When the steps cannot be predicted in advance, the inputs are messy or unstructured, or the decisions need judgement that a rule set cannot capture. For predictable processes, a workflow is the safer default.
Related
Workflow vs Agent
A path fixed in code, compared with a model that chooses its own steps.
Agentic AI vs Generative AI
Content from a prompt, compared with a model that plans, calls tools and finishes the task.
Agent harness
The loop, tools, context, permissions and state that sit around the model.
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
No-code agent builder
Build an agent by configuring it in a visual tool instead of writing code.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.