Skip to main content
Agents are specialist AI personas that you can invoke during development. Each agent has a defined identity, area of expertise, and operating principles. They behave differently from skills: skills carry process instructions that fire automatically, while agents carry personas that you invoke deliberately for their particular expertise.

​
Agents vs. skills

​
How to invoke an agent

Reference an agent by name in your message to Antigravity:
Some agents are invoked automatically by commands — for example, /review uses the code-reviewer agent, and code-reviewer can delegate to security-engineer for security-specific review.

​
All agents

​
The review verdict system

The code-reviewer agent renders one of three verdicts at the end of every review:
  • APPROVED — implementation meets the plan and quality bar; ready to merge
  • APPROVED WITH SUGGESTIONS — implementation is sound but has non-blocking improvements worth addressing
  • CHANGES REQUIRED — critical or important issues must be fixed before merging
A review is only complete when all 6 review areas have been covered and a verdict has been rendered.

​
Customizing agents

Agent files live in ~/.gemini/antigravity/agents/. Each file is a plain Markdown file with a YAML frontmatter block (name, description) and a system prompt body. You can edit any agent file to adjust its persona, focus areas, or output format to suit your project.

Code reviewer

Reviews implementation against the plan across 6 structured areas and renders a verdict.

Backend architect

Designs reliable, scalable server-side systems: APIs, databases, and service boundaries.

Frontend architect

Builds maintainable, performant UIs: component design, state management, and accessibility.

Security engineer

Identifies vulnerabilities, models threats, and designs secure systems using OWASP.

Deep research

Gathers multi-source research with explicit source quality ranking and confidence levels.

System architect

Holistic system design across service boundaries, distributed systems, and migrations.

Socratic mentor

Guides learning through strategic questioning rather than direct answers.

Repo index

Compresses repository context at session start for token-efficient subsequent work.