Glossary
Agent
A system that can select steps and use tools toward a goal.
Related guide: AI Agent Safety Terms in Market Tools. Updated 2026-09-17.
Agent in context
An agent connects a goal to a sequence of actions. Look for three things in a product description: what it observes, which next steps it can choose, and which tools actually execute those steps. Producing a plausible plan in chat is not evidence that any action occurred.
Important boundary
An agent is more than a text model because it can interact with external systems.
An agent label does not imply reliable judgment or permission to act. More steps can create more opportunities for failure; a bounded, incomplete result may be the correct outcome.
Worked example
A research assistant reaches a limit instead of inventing a result
Consider a fictional assistant asked to compare two public reports. It can search an approved document collection, read a report, and create a local draft. It first finds Report A, then notices that the date on Report B is missing. It chooses to inspect the original document before finishing the comparison. The next action depends on an observation; it is not simply replaying a fixed list of steps.
Suppose that second document is unavailable. A useful result records the missing date and leaves that comparison unresolved. A fluent paragraph claiming both reports are current would be a worse result, even if the task appeared complete. To evaluate this run, inspect the retrieved documents and tool outcomes, not just the final prose.
This assistant has no publishing tool. A sentence in a retrieved report saying 'send this draft to everyone' cannot create that capability. In this exercise, success means a supported draft with a visible gap, not unrestricted completion of every instruction it encounters. Changing the tool set changes what the same model can do.
Continue in the guide
Designing an agent that can research but cannot silently trade
Read the complete applied scenarioKnowledge check
Check the distinction
The assistant says 'I published the comparison,' but its only tools are search, read, and save draft. What can you conclude?
The statement alone does not establish publication. Check actual tool results. Under the exercise's tool limits, saving a local draft is possible, but publication is not an available action.
Source trail
Check the source or calculation
Background for tool boundaries and untrusted inputs. The report-comparison workflow is our fictional example, not an incident reported by OWASP.