After two days of hands-on testing, a developer open-sourced carbon-neon-orchestrator, a local plugin for Antigravity 2.0 that assigns architecture to Gemini 4 Carbon and parallel execution to Neon using four hard thresholds, two rounds of circuit breaking, and independent worktrees. We see the competitive focus of AI coding tools shifting from “Which model is stronger?” to “Who can allocate tasks more reliably, control costs, and prevent parallel-development failures?”

What this is

carbon-neon-orchestrator is an open-source plugin for Google Antigravity 2.0. Antigravity is Google’s AI-assisted development environment; Carbon and Neon represent different capability and speed tiers in the Gemini 4 family. The plugin uses a pre-invocation hook—a set of rules triggered automatically before a model executes—to identify the current model. Only Carbon assumes overall control: it breaks down complex architecture, reviews code, and then delegates searches, boilerplate, tests, and similar tasks to Neon for parallel execution.

It does not overcomplicate every simple request. When the path is clear, only one file needs changing, the change is no more than 20 lines, and no build or test is required, Carbon handles the task directly. More involved tasks pass the original image path and existing conversation context to execution nodes. When two Neon instances modify code simultaneously, they use independent worktrees, and Carbon checks the differences before merging. If Neon fails to solve the same issue for two consecutive rounds, Carbon halts parallel execution and takes over itself.

Industry view

The notable design choice is that it turns “model division of labor” from manual selection in a dropdown into automatic task-based routing. Using a top-tier model for decisions and cheaper or faster models for execution can reduce waiting time and model-call costs. For enterprises, the fact that the plugin runs locally and requires no restart also lowers the barrier to experimentation.

That said, this remains an individual developer’s two-day experiment, not production validation at scale. The article provides no comparative data showing whether parameters such as the 20-line threshold and two-round circuit breaker work for other codebases and task types. Independent worktrees also address only part of the risk of code conflicts. A larger question is what happens if more tools depend on a particular client’s plugin hooks and model enumeration: once vendors change their underlying interfaces, plugin-maintenance costs could rise rapidly. Fine-grained orchestration is a practical direction, but this implementation is not yet a universal standard.

Impact on regular people

For enterprise IT: In the past, testers had to manually decide when to activate an expensive model. Now, tasks can be routed automatically according to complexity. Even so, teams should verify code isolation, data permissions, and model-call costs before procurement or deployment rather than relying on demo results alone.

For individual professionals: AI coding assistants may spend less time stuck on complex designs and complete repetitive work faster. The capability worth building will increasingly be defining requirements, validating results, and managing collaboration across tools rather than writing specific code.

For the consumer market: This kind of orchestration may enter IDEs, productivity software, and website-development services more quickly. End users may never need to understand the difference between Carbon and Neon, but they will indirectly benefit from faster delivery and an even quicker pace of software feature updates.