What an AI agent actually is — Devin, Cursor Agent and Claude Code in practice
Agent has become a word vendors apply to anything that calls a function. A more useful definition: an agent plans a multi-step task, takes actions in the world, observes the results, and adjusts — without a human approving each step.
AI Tools Hub Editorial · published · updated
A working definition
- It plans — decomposes a goal into steps rather than answering a prompt
- It acts — edits files, runs commands, calls APIs
- It observes — reads the result, including failures
- It adjusts — changes the plan based on what happened
A chat model that calls a weather API has one of those four. That is a tool call, not an agent. The distinction matters because the failure modes are completely different.
Where the current tools sit
| Tool | Autonomy | What it is good at |
|---|---|---|
| Cursor Composer | Supervised | Multi-file edits you review as they land |
| Claude Code | Semi-autonomous | Running a whole task including tests and git |
| Devin | Autonomous | Taking a ticket to a pull request with minimal intervention |
What agents are actually good at today
Well-specified, close-to-routine work: migrations, mechanical refactors, adding tests to existing code, fixing a bug with a clear reproduction. The common property is that success is verifiable — the tests pass or they do not.
What they are not good at
Anything where success is a judgement call. Architecture, naming, deciding what to build, work that depends on institutional context the agent cannot see. Handing those over produces output that looks finished and is subtly wrong, which is the most expensive kind of wrong.
Will agents replace developers?
They have changed the shape of the work — more reviewing, less typing. The part they do not touch is deciding what should exist and why, which was always most of the job.
How much supervision do they need?
Today, enough that someone reads every diff before it merges. Teams that skip that step find out why within a sprint.
Related
More articles
Opinion
Why AI rollouts fail in Japanese companies — four structural reasons
For business
Enterprise AI rollout checklist — SOC 2, ISO 27001, DPAs and personal data
For business
Japanese-developed LLMs for business — ELYZA, Sakana AI, NTT tsuzumi and PLaMo
Comparison
AI coding tools compared — Cursor, Copilot, Claude Code, Windsurf and Codeium