AI agents are one of those terms that means different things to different people. Vendors use it to sell. journalists use it to sensationalise. And underneath all that noise, there's a genuinely useful and transformative idea. Let me cut through it.
What is an AI agent?
At the simplest level, an AI agent is a system that uses an LLM to decide which actions to take, and then takes them — often in a loop, until it achieves a goal. Unlike a traditional chatbot that answers one question and stops, an agent can plan a sequence of steps, use tools, and adapt based on what it finds.
Think of it as giving an LLM hands. It can not only reason and respond — it can browse the web, read documents, send emails, update a database, or call an API. The LLM is the brain; the tools are the hands.
What can agents actually do?
More than you'd think, and less than the hype suggests. Agents excel at multi-step tasks that require reasoning across different information sources: researching a competitor, synthesising data from multiple reports, drafting a response that needs information from three different documents.
Agents struggle with predictable, rules-based tasks — not because they're too simple, but because the overhead of using an LLM is wasted when a deterministic script would work faster and cheaper. Don't use an agent where a simple automation would suffice.
The practical applications we've built for clients: a contract review agent that reads legal agreements and flags risk clauses, a research agent that monitors competitors and summarises findings daily, a lead qualification agent that scores and routes inbound enquiries, and a content generation agent that drafts personalised outreach emails at scale.
The risks nobody talks about
Agents can hallucinate — confidently produce wrong information. They can take unexpected paths. They can call the wrong tool or use a tool incorrectly. Production agent systems need guardrails, error handling, human oversight for consequential actions, and careful monitoring. Building an agent that works in a demo is easy. Building one that's reliable enough for production requires engineering discipline.