AI agents, explained: what they do and where they still break
An agent is a language model that can take actions in a loop. That one change brings new power, and new ways to fail.
Team Inspiropedia
Key points
- An agent repeats a cycle: decide, use a tool, read the result, decide again.
- Real uses today are narrow: coding help, support triage, research and data tasks.
- The main risks are compounding errors, prompt injection and giving an agent too much access.
A chatbot answers. An agent acts. Technically, an AI agent is a language model placed in a loop with access to tools, a web search, a database query, a code editor, a calendar. The model reads the goal, picks a tool, sees the result and decides what to do next, until it judges the job finished or hits a limit.
What that looks like in practice
- Coding assistants that read a codebase, run tests and propose changes.
- Support tools that look up an order, check a policy and draft a reply for a person to approve.
- Research helpers that search, read sources and compile notes with links.
These work best when the task is well defined, the tools are few and the result can be checked.
Where agents still fail
- Compounding errors. A small mistake early in a long chain can carry through every later step.
- Prompt injection. Text the agent reads, a web page, an email, can contain instructions that try to hijack it. Any agent that reads untrusted content and can take actions needs defences.
- Over-broad access. An agent that can send email or move money can do damage if it is wrong or manipulated.
- Cost and speed. Every loop is another model call.
Sensible guardrails
Give an agent the least access it needs. Require human approval before anything irreversible, such as payments, deletions or messages to outsiders. Log every action so it can be reviewed. And test on real examples before trusting it with real work.
The useful question is not “can it act?” but “what is the worst it could do, and who checks?”