In the span of about two weeks, every major AI lab put a coding agent on the table. Anthropic shipped dynamic workflows in Claude Code with the Opus 4.8 release, fanning a single task across up to a thousand subagents. OpenAI extended Codex past developers entirely, adding Sites, annotations, and enterprise plugins on top of its multi-day Goal Mode. Microsoft used Build to ship MAI-Code-1-Flash, a 5-billion-parameter model priced to undercut the field. And xAI rolled Grok Build into beta — a terminal agent with parallel subagents, worktree support, and ACP, which is to say a near-clone of the others’ feature lists.
Four labs, one product category, two weeks. When the whole industry converges on the same thing at the same time, the convergence is the signal, not the individual launches.
They are all building the same shape
Strip the branding and the four tools rhyme. A terminal-resident agent. Parallel subagents working a problem at once. Git worktrees so those subagents do not collide. A protocol — MCP, ACP — for reaching tools and context outside the model. Long-horizon modes that run for hours or days without a human in the loop.
That convergence is not coincidence or copying so much as the category settling into its natural form. Two years of experimentation — chat sidebars, inline autocomplete, IDE plugins — sorted out what actually moves the needle on real engineering work, and the answer turned out to be an autonomous agent that lives where the code lives and can hold a whole task in its head. Once that shape was proven, every lab with a frontier model had to ship its version or cede the surface. The features stopped being differentiators the moment all four had them.
Why coding, and not something else
Of all the things a frontier model can do, why is this the battleground? Because coding is where AI’s value is most measurable and most verifiable. The output is text, which models are native to. The work has a built-in grader — the code compiles or it doesn’t, the tests pass or they don’t — so an agent can check itself in a loop without a human scoring every step. And the economic value per task is high and legible: an engineer’s time has a price, and shipping more with the same headcount shows up directly in a budget.
There is a stickier reason underneath. A coding agent is where a developer’s workflow physically lives once they adopt it — their context, their habits, their muscle memory for how to prompt and correct it. That is the kind of lock-in the labs have been unable to build at the chat layer, where switching providers costs nothing but a new browser tab. Whoever owns the agent your team works in all day owns a dependency that is genuinely painful to unwind. Every lab understands this, which is exactly why none of them can afford to sit the category out.
The differentiator moved underneath the agent
Here is the consequence that matters for anyone choosing a tool. If the agent’s shape has commoditized — and four near-identical feature lists in two weeks says it has — then the agent is no longer where the competition is decided. What is left underneath is model quality, price, and how well the thing plugs into the systems you already run.
Microsoft’s move is the clearest read on this. MAI-Code-1-Flash is not trying to be the smartest agent; it is a cheap model aimed at the routine majority of coding tasks, sold on the bet that for autocomplete and small edits you do not need a frontier model and should not pay for one. That is what a maturing market looks like: the premium tier keeps the hard problems, and a price-competitive tier takes the volume. The agent becomes a thin client over a model you can swap. When the interface stops being the moat, cost discipline becomes the strategy — the same pattern that has played out in every wave of AI adoption past the hype.
What to do if you are picking one
The instinct in a crowded market is to wait for a winner. That is the wrong move here, because the productivity gains are real now and the shakeout will take quarters. The right move is to adopt, but adopt for portability.
Favor the parts of your setup that survive a vendor switch. Protocols like MCP and ACP exist precisely so your tool integrations are not married to one lab’s agent; lean on them. Keep your prompts, your task definitions, and your review process in your own repo, not locked in a vendor’s proprietary format. Treat the agent as replaceable and the workflow around it as the asset. If a cheaper model clears the bar for your common case next quarter — and on current trajectory one will — you want to be able to take it without rebuilding everything.
The uncomfortable part is that the labs are counting on you not doing this. The whole point of putting the agent where your work lives is that switching becomes too annoying to bother with. The crowding is good for buyers right now — four vendors competing on price and capability in the same quarter is the best leverage you will get. It stays good only as long as you refuse to let any one of them become load-bearing in a way you cannot undo.