Is Claude Code an agent?
Yes — Anthropic calls it an "agentic coding tool", and it passes the operational test rather than just the marketing one: given a goal it decides which files to open, edits several, runs your commands, reads the output and tries again, without a prompt per step. It also spawns subagents that split a task in parallel. What it isn't is autonomous — it asks permission before acting.
The label is the vendor's; the behaviour is what settles it.
Anthropic's own definition opens with "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools". Labels are cheap, so the more interesting evidence is the shape of what it does: the docs' own example prompt is "write tests for the auth module, run them, and fix any failures" — one instruction covering discovery, authorship, execution and repair.
Two features only an agent needs. Subagents: several Claude Code agents work on different parts of a task while a lead coordinates and merges. The Agent SDK: build your own agents on the same tools and orchestration. Neither makes sense for a completion tool.
And a limit that matters more than the label. It is not autonomous: by default it asks before changing files, and it asked us for approval before running a script and before searching the web. Agency here means deciding what to do, with a human still gating the doing.
What makes it agentic.
A completion tool answers the question in front of it. An agent chooses the next question. That's the whole distinction, and it's visible the first time you give Claude Code a goal instead of an instruction: it goes looking, forms a plan, and comes back having changed several files.
The feedback loop is the second half. Running your tests and reading the failure is what separates an agent from a very good code generator — it can tell whether it succeeded.
Worth being precise, because "agent" makes people imagine something unsupervised. In default mode it proposes and asks. We watched it decline to search the web until the tool was approved, and offer to write a script rather than silently running one. There are modes that loosen this, up to a flag whose name is a warning label.
So: autonomous in its reasoning, supervised in its effects. If you want it unsupervised in CI, that's a deliberate configuration rather than the default.
Beyond a single session, it can spawn subagents that take different parts of a task simultaneously, with a lead agent assigning work and merging results, and background sessions you can watch from one screen. The Agent SDK exposes the same machinery for building your own.
That's the part where the word stops being a marketing adjective and starts describing an architecture.
Other agentic coding tools.
The category now has several credible entries; they differ on where the agent runs and what it costs rather than on whether they qualify.
Frequently asked.
Quick follow-ups people search after this question.