スペック駆動開発の進化：ConductorがAntigravityをサポートEvolving Spec-Driven Development: Conductor Now Supports Antigravity

Google's Conductor now supports Antigravity, a change that extends spec-driven development workflows to AI agents built on the Gemini model family. The update matters because it moves agentic coding further away from ad hoc prompting and toward a more structured, specification-first process, where a written spec becomes the shared source of truth that both humans and agents work from.

Spec-driven development is the central idea here. Rather than typing a loosely worded instruction and hoping an AI produces the right code, developers first author a structured specification that captures requirements, intended behavior, constraints, and often a breakdown of implementation tasks. That document then drives the actual work. The approach aims to reduce ambiguity, make agent output more predictable, and give teams an artifact they can review, version, and reason about before any code is generated. It also creates a cleaner audit trail, since the reasoning behind a change is written down instead of buried in a transient chat history.

Conductor, in this context, appears to act as an orchestration layer that coordinates how agents execute against those specifications. By adding support for Antigravity, it lets developers route spec-driven tasks into Google's agent-oriented development environment. Antigravity is Google's agent-first platform for software development, designed so that AI agents can plan, write, and iterate on code with a degree of autonomy rather than functioning purely as inline autocomplete. Pairing an orchestration tool with that environment is likely intended to make multi-step, multi-agent work more manageable, so that a single specification can be decomposed into tasks and carried out with clearer boundaries.

The Gemini foundation is what powers the reasoning underneath. Because the agents are built on Gemini, the integration benefits from the model's capabilities in code understanding, long-context reading, and tool use. In a spec-driven flow, long context is particularly relevant: the model needs to hold the specification, the relevant portions of the codebase, and the current task in view at once to produce changes that stay faithful to the stated intent. Keeping the spec as an explicit input, rather than relying on the model to infer requirements, tends to improve consistency across a series of related edits.

It helps to place this within the broader industry direction. Spec-driven development has gained momentum across the tooling landscape, with efforts such as GitHub's Spec Kit and Amazon's Kiro promoting similar ideas: define specifications and plans first, then let agents implement against them. The common thread is a recognition that fully autonomous coding agents can drift or hallucinate requirements when given only vague prompts, and that a formalized specification acts as a guardrail. Google's move to connect Conductor with Antigravity fits this pattern and signals continued investment in structured agent workflows rather than one-shot generation.

For teams considering the approach, a few prerequisite concepts are worth understanding. Effective spec-driven work assumes some discipline in writing specifications that are precise enough to be actionable but not so rigid that they prevent iteration. It also assumes a workflow where agents can be observed and corrected, since even well-specified tasks may require human review at checkpoints. Familiarity with agent orchestration, where multiple steps or multiple agents are sequenced and coordinated, becomes increasingly important as projects grow beyond a single file or function.

The practical impact of this integration is likely to be felt most by developers already experimenting with agentic tooling who want more repeatable results. By combining a specification-first methodology, an orchestration tool, and an agent-native environment running on Gemini, the setup aims to make structured development more powerful without abandoning the flexibility that made AI coding assistants appealing in the first place.

As with any fast-moving area, the real test will be adoption and reliability in day-to-day use. Structured workflows can add overhead, and their value depends on whether the resulting agent output is accurate enough to justify writing detailed specs up front. Still, the direction is clear: the ecosystem appears to be converging on specifications as the connective tissue between human intent and autonomous agents, and this update is another step along that path.