Google's Gemini 4 Carbon makes multi-agent coding more practical

Google's latest AI launch is as much about how developers organise AI work as it is about a new model. On 10 November 2026, the company made Gemini 4 Carbon available to paying developers and enterprise customers, while previewing a feature in its Antigravity coding environment called /swarm.

The idea is straightforward: instead of asking one AI assistant to plan, write, test and review a complicated change, a developer can coordinate several agents working on different parts of the job.

One prompt, several workers

Google describes /swarm as a way to launch a coordinated team of agents from a single prompt. Its proposed setup uses Carbon for demanding work such as architecture and auditing, alongside faster models working in separate environments. Other agents can check the changes.

This is different from simply opening multiple chat windows. The interesting part is the coordination: deciding which tasks can run in parallel, keeping changes isolated and bringing the results back together.

For a real software project, that might mean one agent investigating a bug, another proposing a fix and a third testing whether the fix breaks something else. That's an illustrative workflow, not a claim that Antigravity can already perform every part of it reliably.

Why Google has two flagship models

Carbon follows Gemini 4 Argon, which Google says remains the better fit for most software engineering and knowledge work. Carbon is intended for harder, longer-running tasks that justify more reasoning and verification.

That distinction matters because the most capable model is not automatically the most economical choice for every step. A team of agents could use an expensive model for planning and difficult decisions, then delegate routine tasks to cheaper ones.

Google's introductory Carbon API price is $2 per million input tokens and $10 per million output tokens. The company says those rates double after 30 November 2026. That makes cost monitoring especially relevant for workflows in which several agents make repeated model calls.

What the launch does not prove

Google reports strong benchmark results for Carbon, but benchmark rankings are not a guarantee that an agent will understand a particular codebase or make safe changes in production. The Verge's coverage also notes that the company's choice of benchmarks deserves scrutiny.

Multi-agent systems introduce their own problems. Parallel work can conflict, duplicated effort can waste tokens and a flawed plan can spread errors across several agents. More agents also mean more permissions and actions to audit.

Developers should still use isolated branches or workspaces, automated tests, code review and explicit approval for consequential changes. An agent that confidently reports success is not a substitute for a passing test suite.

The bigger shift

The useful question is no longer just which model writes the best code. It's how an AI development environment divides work, checks results and gives the developer enough visibility to stay in control.

Carbon and Antigravity's swarm preview are another step towards that model of software development. The promise is faster progress on complex tasks. The test will be whether coordinated agents produce fewer mistakes and less review work than one well-managed assistant.