Some weeks in tech are quiet. This wasn't one of them. In the space of seven days, OpenAI, GitHub, Google and Anthropic all shipped or showed coding agents. If you blinked, you missed the moment AI coding moved from "autocomplete with ambitions" to "a teammate you hand an issue to".

The headline on Thursday was Claude 4. But the real story is the week as a whole, so let's look at what actually happened, and what I think it means for developers who ship real software.

The week, in order

Four companies, one direction. That's not a coincidence of calendars. It's the industry agreeing on what comes after chat.

What Anthropic actually shipped

Two new models: Claude Opus 4, the larger one, and Claude Sonnet 4, the everyday workhorse. Anthropic's own announcement focuses heavily on coding and on long, multi-step agent work, rather than on chat. You can read it in full on the Anthropic news page. I won't repeat benchmark numbers here; they're useful for comparing models, but they tell you little about how a model feels on your own codebase. That you only learn by using it.

For me, the bigger news was Claude Code becoming generally available. It's been in research preview since February, and I've been using it daily. GA brings integrations with VS Code and JetBrains IDEs, so the agent's edits show up right where you already work, instead of only in the terminal.

The pattern behind all four launches

Strip away the branding and the launches share a shape:

• The agent works on a task, not a line. You describe an outcome ("fix this bug", "add this endpoint"), not the next few characters.
• It works in a real environment. It reads the repo, runs commands and tests, and checks its own output.
• It hands back reviewable work. A diff or a pull request, not a wall of text to copy and paste.
• It fits existing tools. GitHub issues, the IDE, the terminal. Not a separate website you have to switch to.

Some run in the cloud, in the background, while you do other things. Some run locally, next to you, asking before risky steps. Both styles will stay around, because they fit different jobs.

What this means if you write code for a living

1. Reviewing becomes a bigger part of the job

If an agent can produce a pull request from an issue, the scarce skill is no longer typing the code. It's reading a diff and knowing whether it's correct, safe and maintainable. Code review was always important. It's about to become central.

2. Your tests are now your guardrails

An agent checks its work by running your test suite. If the suite is thin, the agent's "all green" means very little. Teams with good tests will get far more out of these tools than teams without them. That's the least glamorous, most important takeaway of the week.

3. Clear issues get better results

"Fix the bug in payments" gives an agent the same problem it gives a new colleague. Issues with steps to reproduce, the expected behaviour and acceptance criteria were always good practice. Now they directly improve what comes back.

4. Project instructions pay off

Every one of these tools works better when it knows your conventions: how to run the tests, which patterns to follow, what never to touch. In Claude Code that's a CLAUDE.md file. Writing one is the cheapest improvement you can make, whichever agent you use.

5. The fundamentals matter more, not less

To review an agent's database migration, you have to understand migrations. To spot a security hole in its auth code, you have to understand auth. Agents make experienced engineers faster. They don't replace the experience.

A healthy dose of caution

Announcement week always looks better than month three. Agents still make confident mistakes, still need clear boundaries on what they can run, and still need someone to read every diff before it reaches production. I'll be trying the new models on real tasks over the coming weeks, with the same rule as always: nothing merges without a human review.

Which of these are you trying first: Codex, the Copilot agent, Jules or Claude Code? And what kind of task would you trust an agent with on day one?