What is it?
"Agentic" describes systems that use models to take multi-step, tool-using action. Within that, there is a spectrum:
- Workflow: code defines the sequence; the model performs bounded sub-tasks. Predictable and testable.
- Agent: the model chooses steps dynamically in a loop (What are AI Agents?). Flexible but less predictable.
Real systems sit between these. An agentic workflow is a designed process in which model calls, tools, checks and sometimes humans are combined deliberately.
Why does it matter?
Reliability compounds. If each of five steps is 90% reliable, the whole chain succeeds only about 59% of the time (0.9⁵) — an illustration, not a measurement. Designing the path lets you place validation after risky steps, retry cheaply, use smaller models for easy steps, and test each stage separately. It also keeps cost and latency understandable.
How does it work?
Reusable patterns:
- Prompt chaining: break a task into sequential model calls, optionally with programmatic checks ("gates") between them. Outline → check the outline → draft each section → assemble.
- Routing: classify the input and send it down the right path or to the right prompt/model. Billing question → billing prompt; bug report → triage prompt.
- Parallelisation: run independent calls at once — different sections, or several attempts and vote — then merge.
- Orchestrator–workers: a model decides the sub-tasks at run time, workers execute them (Multi-Agent Systems Explained).
- Evaluator–optimiser: one call produces, another critiques against criteria, loop until acceptable or limit reached.
- Human-in-the-loop checkpoints: approval before irreversible or high-stakes steps.
- Fallbacks: when confidence is low or validation fails, escalate to a person or a safer default.
Implementation tips: keep state explicit (a JSON object passed between steps); validate each model output against a schema; log every step; set budgets (steps, tokens, time).
Example
Invoice processing:
- Extract fields from the document with a model (schema-validated).
- Check in code: totals add up, vendor exists, dates valid.
- Route: if checks pass and amount is below a threshold → create a draft entry; otherwise → flag for human review with the reasons.
- Human approves before posting.
The model handles the messy part (reading varied layouts); everything else is ordinary, testable code. There is no open-ended autonomy because the task does not need any.
When should I use it?
Default to the simplest thing that works: a single well-prompted call; then a workflow; then, only if the path truly cannot be predetermined, an agent. Choose workflows when the steps are known, accuracy and auditability matter, or cost must be controlled. Choose agents for open-ended problems where the number and order of steps depend on what is discovered.
Common mistakes
- Jumping to full autonomy when a fixed pipeline would do.
- No validation gates, so errors flow through to the end.
- Opaque state, making debugging impossible.
- Unbounded retries and loops.
- Evaluating only end results, not each step (How to Evaluate AI Systems and Outputs).
- Skipping the human checkpoint where the consequences justify one.