Back to Blog
AI

Google Conductor Plugin: Spec-Driven Coding Agents Move Beyond One CLI

Medianeth Team
July 22, 2026
8 minutes read

The next wave of AI coding tools is not just about smarter models.

It is about memory, procedure, and proof.

Google's July 16 Conductor update is useful for that reason. Conductor started as a Gemini CLI extension for spec-driven development. Google now says it has evolved into the Conductor Plugin, a portable package that can work across coding-agent environments such as Antigravity CLI and Claude Code.

That sounds like a tooling detail. It is bigger than that.

For teams trying to use coding agents on real software, the hard part is rarely "can the model write code?" The hard part is keeping the work tied to product goals, architecture decisions, plans, and reviewable artifacts after the chat scrolls away.

Conductor is Google's answer to that problem: make the repo, not the chat thread, the source of truth.

What happened

On July 16, 2026, Google announced that Conductor is moving from a Gemini CLI extension into the Conductor Plugin.

Google says plugins can bundle skills, rules, MCP servers, and hooks in one package. In practice, that means Conductor can carry the workflow protocol around with the coding agent instead of living as a single-CLI extension.

The new positioning is simple:

  • Conductor still uses persistent Markdown artifacts such as spec.md and plan.md.
  • The workflow is now more conversational, so developers do not have to drive everything through strict command sequences.
  • The plugin is compatible with Antigravity CLI.
  • The GitHub repository describes Conductor as a plugin for AI coding agents, including Antigravity and Claude Code.
  • Google's codelab shows install paths for both Antigravity CLI and Claude Code.

The practical shift is from "use this extension inside Gemini CLI" to "bring this spec-driven workflow into the agent tool you are already using."

Why people are talking about it

AI coding agents are getting better at executing.

That creates a new problem: execution without shared project state can get messy fast.

If an agent only has the current prompt and a partial view of the repo, it may produce code that looks plausible but drifts from the product goal, ignores architectural constraints, repeats a rejected approach, or implements a feature before the team agrees on what the feature is supposed to do.

Spec-driven development tries to fix that by moving intent into durable files.

Instead of asking an agent to "build auth" from a chat prompt, the team keeps the goal, constraints, plan, and progress in version-controlled Markdown. The agent reads those artifacts, updates them as work moves, and uses them to guide implementation.

That is less magical than vibe coding. It is also more useful when multiple humans, multiple agents, and multiple sessions touch the same codebase.

What is confirmed

Google's announcement confirms that Conductor now supports Antigravity through the Conductor Plugin and that the plugin keeps the core spec-driven development artifacts.

The same announcement says Conductor can dynamically generate context, specs, and plans as the developer discusses requirements with the assistant. Google also says the plugin can update context or check off completed plan tasks in the background.

The GitHub repository confirms the plugin framing and describes the lifecycle as:

  1. Context.
  2. Spec and plan.
  3. Implement.

The repository also lists installation instructions for Antigravity and Claude Code, plus features such as planning before building, maintaining context, iterating safely, project-level context, brownfield initialization, and a git-aware revert flow.

Google's codelab confirms a hands-on workflow: install the plugin, initialize a project, generate context files, create a track, review a generated spec.md and plan.md, then implement against those artifacts.

What is still unclear

This update does not prove Conductor will make every coding agent reliable.

It also does not prove that a Conductor-managed plan is automatically the right plan.

The open questions are the ones teams should test in their own repos:

  • How well does the plugin preserve intent across long projects?
  • Does the generated plan match the team's actual architecture?
  • Can developers review and correct the artifacts without fighting the tool?
  • How cleanly does the workflow behave in messy brownfield codebases?
  • What happens when different agents or humans disagree with the saved plan?
  • Does the added planning overhead pay for itself on small tasks?

Those are not reasons to ignore Conductor. They are reasons to pilot it on a bounded feature instead of immediately turning it into company-wide process.

Why it matters for businesses

Most companies do not need more AI demos.

They need repeatable delivery.

That is where spec-driven coding workflows become interesting. They turn AI-assisted development from a private chat habit into something closer to an auditable engineering process.

For an agency, product team, or operations-heavy business, that matters because the expensive failures are usually not syntax errors. The expensive failures are misread requirements, hidden assumptions, missing acceptance criteria, and agents making local decisions that conflict with the business goal.

Conductor's promise is not "the agent will be perfect." The safer claim is this:

If the specification, plan, and progress live in the repo, humans can inspect what the agent thinks it is building before and after the code changes.

That gives teams a better control surface.

Where Conductor fits

Conductor is most useful when the work has enough shape to deserve a plan.

Good candidates:

  • New features with multiple steps.
  • Brownfield changes where the existing architecture matters.
  • Agent-assisted work that spans more than one session.
  • Client work where decisions need to be explained later.
  • Features with acceptance criteria, constraints, or review gates.

Weak candidates:

  • Tiny copy edits.
  • One-line bug fixes.
  • Exploratory spikes where the goal is still unclear.
  • Tasks where a human has not decided the product behavior yet.

The clean mental model is: use Conductor when forgetting context would be expensive.

A practical adoption checklist

If your team wants to try Conductor, keep the first pass boring.

1. Pick one bounded feature

Choose a feature with clear business value and enough complexity to test the workflow. Avoid a production incident, a huge migration, or a speculative research spike.

2. Review the generated artifacts before code

Do not let the agent jump straight from conversation to implementation. Read the context, spec, and plan first. Correct wrong assumptions while the mistake is still cheap.

3. Treat Markdown as the contract

If the plan changes, update the plan. If the acceptance criteria change, update the spec. The workflow only helps if the durable files stay closer to truth than the chat log.

4. Keep human approval on risky steps

Conductor can organize the work, but it should not silently broaden permissions, touch production data, deploy infrastructure, or rewrite security-sensitive code without review.

5. Compare against your current process

Measure whether the workflow reduces rework, improves review quality, or makes handoff easier. If it only adds ritual, shrink the process.

What not to overclaim

Conductor is not a substitute for product judgment.

It does not remove the need for code review, tests, threat modeling, or clear ownership. It does not prove that Antigravity, Claude Code, Gemini CLI, or any other agent will make the right call in your codebase.

The confirmed improvement is narrower and more useful: Google is turning a spec-driven coding workflow into a portable plugin with persistent repo-local artifacts and conversational interaction.

That is enough to care about.

The future of AI coding will not be one perfect agent guessing correctly from a paragraph. It will be agents that can work from shared context, leave evidence behind, and let humans intervene before the wrong thing becomes expensive.

What to do next

If your team already uses coding agents, try this on one medium-sized feature:

  1. Write the product goal in plain English.
  2. Generate or draft a spec and plan.
  3. Review the plan before implementation.
  4. Let the agent implement one phase.
  5. Compare the diff against the saved acceptance criteria.

If the saved artifacts make review easier, keep going. If they expose wrong assumptions, that is still a win. You found the problem before it shipped.

That is the quiet value of spec-driven development: not more AI output, but fewer invisible misunderstandings.

Sources checked

No trend-only sources were used for this article. The recommendations are Medianeth's interpretation of the confirmed Google and GitHub documentation, not a claim that Conductor alone makes AI-generated code safe.

Note: This article was prepared with AI assistance and checked against primary sources before publication.

Your Next Project, Delivered in 8–12 Weeks

Tell us what you're building. We'll show you the fastest path to a production-ready launch.

Get My Free Proposal