"Agent" is currently doing the work "cloud" did in 2010: meaning everything, clarifying nothing. Vendors call a chatbot with a system prompt an agent. Skeptics call the whole idea vaporware. Both are wrong in useful ways.
Here's the definition I actually build with.
The loop is the thing
An AI agent is a system where a model runs a loop: plan → act → observe → adjust. It decides on a step, uses a tool (runs code, searches, calls an API, edits a file), looks at what happened, and decides the next step — repeating until it hits the goal or gives up.
That's it. No consciousness, no autonomy in the sci-fi sense. A loop, tools, and a goal.
The loop matters because it's the difference between answering and doing. A chat model asked about a failing test explains what might be wrong. An agent runs the test, reads the failure, edits the code, runs it again — and hands you a passing suite.
Tools define the ceiling
An agent with no tools is a chat window. An agent with your terminal, your codebase, and your issue tracker is a junior engineer. An agent with your production database is a liability.
This is the design insight most agent talk skips: capability and risk are both set by tool access, not intelligence. The interesting engineering work in agents is deciding what they can touch, what requires approval, and what's off-limits — the same access-control thinking we've always done, applied to a new kind of actor.
When agents beat assistants
Agents earn their complexity when three things are true:
- The task is multi-step — real work with intermediate results that
inform later steps.
- "Done" is checkable — tests pass, the document matches the schema,
the numbers reconcile. Agents thrive on verifiable goals.
- The cost of iteration is low — steps are safe to retry, or sandboxed
so mistakes are cheap.
Reformatting a document doesn't need an agent. Migrating three hundred call sites, where each edit follows the same checkable pattern? Agent work. Deciding your product strategy? Not agent work — no loop fixes an unverifiable goal.
The failure modes are boring (that's good news)
Agents fail the way flaky software fails, not the way movies predicted: they wander when the goal is vague, they double down on wrong assumptions, they burn budget retrying the unfixable. All of it is manageable with the tools engineering already has — clear specs, sandboxes, step limits, review gates, and logging.
Which is the real takeaway: agents are not magic and not a scam. They're a new kind of software component with a new failure profile, and the people getting value from them are the ones treating them exactly that seriously.