Core agents
What is Agentic Workflow?
An agentic workflow is a sequence of observable stages an agent runs to finish a job.
Typical stages include input, planning, tool use, verification, and output. Failures and retries are part of the workflow.
Visual
Agentic workflow stages
A typical five-stage workflow with failure and retry paths.
Why it matters
Without a defined workflow, an agent's behavior is a black box. Agentic workflows make each stage — input, planning, tool use, verification, output — explicit and inspectable. This is what separates debuggable systems from ones that "sometimes work."
Workflows also make failure expected. A retry after a failed tool call is part of the workflow, not a bug.
Typical stages
Most agentic workflows follow a pattern: receive the task, plan or decompose it, execute tool calls, observe results, verify the output against criteria, and emit the final artifact. Some workflows add reflection (re-checking the plan) or escalation (asking a human).
The number of stages varies, but the principle is constant: every transition should be logged and inspectable.
Key takeaways
- 1Workflows make agent behavior explicit, inspectable, and debuggable.
- 2Failure and retry are designed stages, not accidents.
- 3Every stage transition should be logged in the trace.
Common mistakes
- ✕Running agents without any workflow definition and hoping the model figures it out.
- ✕Skipping the verification stage because it "slows things down" — it also catches errors.
In practice
In production, workflows are usually defined in code — not in the prompt. The controller advances stages deterministically while the model handles the reasoning within each stage. This separation makes testing and monitoring straightforward.
Related terms
Concept neighborhood
Terms linked from Agentic Workflow in the glossary graph.
- Agentic Workflow
- Agent
- Planning
- Verification
FAQ
- Can a workflow have dynamic stages?
- Yes. The model can decide to add a verification step or skip one if the task is simple. The key is that each decision is logged.