An agentic workflow is a process in which AI carries out or chooses multiple steps towards a goal, rather than producing a single response. The term covers a range: at one end, a fixed chain of model calls designed by a developer; at the other, an agent deciding its own path. Choosing the right point on that range is the main design decision in agent engineering.
Key takeaways
- Workflows follow predefined paths; agents choose their own.
- Predictable tasks should use workflows: cheaper, testable and easier to debug.
- Common patterns include chaining, routing, parallelisation, evaluator-optimiser and orchestrator-worker.
- Most reliable systems combine fixed structure with bounded agent freedom.
- Human checkpoints belong wherever mistakes are costly.
An agentic workflow explained in simple terms
A recipe tells you exactly what to do in order: that is a workflow. A chef told “make something with what’s in the fridge” decides everything: that is an agent. Most professional kitchens run recipes for consistency and let the chef improvise only where it adds value. Software should do the same.
Common workflow patterns
| Pattern | How it works | Good for |
|---|---|---|
| Prompt chaining | Output of one call feeds the next | Multi-stage content, extract-then-summarise |
| Routing | A classifier sends input down a specialised path | Support triage, mixed request types |
| Parallelisation | Several calls run at once, results combined | Multiple checks, voting, speed |
| Evaluator-optimiser | One call produces, another critiques, repeat | Drafting, code, translation quality |
| Orchestrator-worker | A lead model plans and delegates subtasks | Complex research, multi-agent work |
| Agent loop | The model decides each step with tools | Open-ended tasks with unknown paths |
A worked example
Handling an incoming support email:
- Routing: classify as billing, technical or sales.
- Retrieval: fetch relevant help articles for that category (RAG).
- Generation: draft a reply grounded in those articles.
- Evaluator: a second call checks the draft against policy rules.
- Branch: if confidence is high and the request is routine, queue for sending; otherwise route to a human.
Steps 1 to 5 are fixed. Only step 2’s search query is model-chosen. This is a workflow with a small agentic component, and it will be far more reliable than handing the whole task to an open-ended agent.
Workflow vs agent: choosing
| Question | If yes, use |
|---|---|
| Are the steps known in advance? | Workflow |
| Does every case follow a similar path? | Workflow |
| Do you need predictable cost and latency? | Workflow |
| Can the number of steps vary widely? | Agent |
| Does the path depend on what is discovered mid-task? | Agent |
| Is exhaustive testing essential? | Workflow |
A widely repeated principle in agent engineering is to use the simplest pattern that solves the problem, adding autonomy only when simpler designs demonstrably fail.
Designing reliable agentic workflows
- Define clear success criteria per step.
- Validate structured outputs between steps.
- Add retries with different prompts for transient failures.
- Cap total steps, time and token spend.
- Place human approval before irreversible actions.
- Log inputs, outputs and decisions at every step.
Limitations
- Errors propagate through chained steps.
- More steps mean more latency and cost.
- Agentic branches are hard to test exhaustively.
- Complex workflows need ongoing maintenance as tools and prompts change.