01
The agent loop, which is the whole idea
A normal chat turn is one shot: you ask, the model produces text, it stops. An agent turn is a loop. The model decides on an action, a tool executes it, the result comes back as new input, and the model decides again. It repeats until the task is finished or it gives up.
This is why Claude Code can do things a chat window cannot. Asked to fix a failing test, it can read the test, read the implementation, make an edit, run the suite, see the failure, and try again. None of those steps require new intelligence; they require the ability to observe consequences.
- Read and search files across a project.
- Run commands and read what they print.
- Edit files directly rather than proposing a patch.
- Loop on the output until the goal is met.
02
Why file access changes the quality, not just the convenience
Pasting code into a chat window means the model sees the fragment you chose. If the bug is caused by something you did not paste, it cannot find it and will confidently theorise about what it can see.
An agent with repository access can go looking. That difference shows up most on problems where the cause is somewhere other than the symptom, which is most real debugging.
It is also the reason to supervise it. A tool that edits files is a tool that can edit files wrongly. Running it inside a git repository with a clean tree is not fussiness, it is the mechanism that makes every change reviewable and reversible.
03
Skills: making a procedure repeatable
A Skill is a folder containing a required SKILL.md file, plus optional scripts and supporting resources. The markdown describes when the skill applies and what to do; the scripts do the deterministic parts.
The point is loading on demand. Rather than stuffing every instruction into a system prompt where it competes for attention, the model pulls in the relevant skill when the task matches, which keeps context focused.
Practically, a Skill is how you stop re-explaining your deployment process, your review checklist or your house style every single time.
04
Claude Cowork, and Record a Skill
Cowork applies the same agent pattern to work that is not software: operations, finance, content. It connects to your data, can be scheduled to run on a recurring basis, and produces documents, decks and spreadsheets rather than commits. Anthropic has reported that more than 90 percent of Cowork usage is non-software work.
The most interesting addition is Record a Skill, which Anthropic shipped in the Cowork desktop app in July 2026. You record your screen while narrating a task, and Claude converts the demonstration into a reusable skill.
That closes a real gap. Writing a good skill by hand requires knowing how to specify a procedure precisely, which is itself a skill. Demonstrating a task is something anyone who does the task can already do.
05
What the loop does not fix
An agent is the same model underneath, so the failure modes described on our ChatGPT page still apply. It can misread a codebase, be confidently wrong about why something breaks, and write a plausible fix that passes the wrong test.
What the loop adds is feedback: an agent that runs the tests finds out it was wrong, where a chat window never does. That is a genuine improvement in reliability and not a replacement for reading what it did.
Frequently asked questions
What is the difference between Claude Code and Claude in the browser?
The browser version sees only what you paste and returns text. Claude Code runs in your terminal with access to your project, so it can read files, run commands and apply edits in a loop. Same model, very different reach, which is why one needs supervision and the other does not.
What exactly is a Claude Skill?
A folder containing a required SKILL.md file plus optional scripts and resources. The markdown says when the skill applies and how to carry it out; the scripts handle deterministic steps. Claude loads it when the task matches rather than carrying every instruction in context all the time.
How is Claude Cowork different from Claude Code?
Same agent idea, different audience. Claude Code targets engineers working in a terminal on a codebase. Cowork targets knowledge work across documents and connected tools, produces decks, documents and spreadsheets, and can run on a schedule. Anthropic reports the large majority of its use is non-software work.
Do I need to code to use Record a Skill?
No, that is the point of it. You record your screen while narrating what you are doing, and Claude turns the demonstration into a reusable skill. It converts knowing how to do a task into a specification, without you having to write the specification.
Is Claude Code safe to let loose on my repository?
Treat it as a capable colleague who is occasionally confidently wrong. Run it inside git with a clean working tree, read the diff before accepting, and keep it away from production credentials. The risk is not malice, it is a plausible-looking change that is wrong in a way you did not check.