Glossary
AI agent
Also: Agentic workflow
A language model that does not only answer but calls tools over several steps, judges intermediate results and continues.
The difference to a simple call is the loop: the model decides which tool it needs, sees the result, decides again. That turns a text generator into something that gets tasks done, and a predictable call into an operation of open length.
That is precisely the central operational property: an agent has no fixed number of steps, no fixed runtime and no fixed bill. Without hard upper limits the price per operation is unbounded, and a loop that gets stuck is the one case in which an AI bill runs away overnight.
The second property is the blast radius. Whatever an agent can do, it can also do wrongly, and it does so on its own, often repeatedly. The question is therefore never "how good is the agent" but "what happens in the worst case, and how do we notice".
There is an honest middle ground that is often enough in practice: a hard wired flow with model calls at defined points. No agent, no loop, but predictable cost, clear error handling and repeatability. A great many tasks that get built as an agent are in truth a workflow with three model steps.
Where a real agent carries its weight is tasks with an unknown sequence of steps: research across several sources, fault finding, tasks whose route only emerges from intermediate results. There the freedom is the point rather than the risk.
With agents, observability is not an extra, it is a precondition. Without a trail covering every step, every tool call and every intermediate result, you can neither trace an error nor demonstrate an improvement.
How you notice it
- There is no upper limit on steps or cost per operation.
- The flow could be written down as a flowchart but is built as an agent.
- Writing tools run without confirmation.
- After misbehaviour it is impossible to reconstruct which steps ran.
Not to be confused with
- Workflow orchestration
- Hard wired steps whose order sits in the code. Predictable in cost and behaviour. For most business processes the better choice.
- Function calling
- A single tool call within one answer. An agent is the loop around it, calling and judging several times.
- RPA
- Automates click paths in user interfaces, deterministic and brittle. An agent works through interfaces and is more flexible but less predictable.
- Chatbot
- Answers but does not act. The transition to an agent begins with the first tool that changes something.
When it fits
- The sequence of steps is not foreseeable and only emerges from intermediate results.
- The task requires research or search across several sources.
- Failed attempts are cheap and without consequence, for example when reading.
When it does not
- With a fixed flow of known steps: then a workflow is the right answer.
- For actions that move money or cannot be undone, without a confirmation stage.
- Without an upper limit on steps, time and cost per operation.
- Without a log of every step.
How to approach it
- Check first whether a workflow is enoughIf the sequence of steps can be written down, it belongs in the code. A deterministic flow with model calls at three points is cheaper, faster and verifiable.
- Cut tools small and along the business domainFew, clearly described tools. Every additional one costs context and raises the probability of a wrong choice.
- Set hard limitsMaximum number of steps, time limit, cost ceiling per operation. On exceeding them, abort and escalate rather than letting it run on.
- Secure writing actionsConfirmation by a person or a second check for anything that cannot be undone. Reading tools free, writing tools deliberate.
- Log every stepInput, chosen tool, parameters, result, decision. Without that trail misbehaviour cannot be reconstructed.
- Make silent failures loudAn agent that has stopped doing anything goes unnoticed. Alert on missing or aborted operations, not only on errors.
Frequently asked
When is an agent worth it instead of a fixed flow?
When the sequence of steps is not known in advance. Research, fault finding and tasks whose route emerges from intermediate results all benefit. Anything that can be written down as a flowchart belongs in a workflow: predictable cost, clear error handling, reproducible behaviour.
How do you keep the cost under control?
With three hard limits: maximum number of steps, time limit and cost ceiling per operation, all in the code and not in the prompt. Plus a measurement of the actual distribution: the mean number of steps per task is less interesting than the upper percentile.
May an agent write on its own?
Reading gladly, writing only with safeguards. Anything that moves money, communicates to the outside world or cannot be undone needs a confirmation stage. At high volume at least thresholds and spot checks, rather than approving every operation individually.
How do you test an agent?
With a fixed set of tasks and known results, as with any other AI system, plus a check of the edge cases: what happens on a tool error, on contradictory intermediate results, on reaching the step limit. Those cases determine behaviour when it matters more than the normal flow does.