Antigravity CLI Agent reliability issue

I am using Antigravity CLI as part of a multi-agent software development workflow.

The issue is not only model capability, but the reliability of the coding agent during complex project execution.

I initially used the 3.6 high model. After encountering repeated issues, I switched to the 3.1 Pro model, but the same types of problems still occurred.

The main problems I observed:

  1. Project instruction following

The agent sometimes fails to consistently follow established project rules and constraints.

Examples:

  • ignoring existing development rules
  • changing files outside the assigned scope
  • making assumptions instead of asking for clarification
  1. Completion status accuracy

The agent often considers a task “completed” when only a prototype or partial implementation is finished.

Examples:

  • code generated but not fully integrated
  • UI exists but API contracts are inconsistent
  • tests pass in a limited environment but production conditions are not verified
  1. Engineering validation gaps

Important engineering aspects are sometimes missed:

  • version control and backup verification
  • commit/hash consistency
  • API/DTO contract consistency
  • security boundaries
  • production permission models
  • failure path testing
  1. Risky autonomous changes

The most serious issue is that incorrect changes can require other coding agents to repair them.

In a real software engineering workflow, an incomplete task is usually less harmful than an autonomous agent making incorrect changes to an existing system.

My expectation:

Antigravity CLI is very useful for rapid prototyping and implementation, but for complex projects it needs stronger:

  • project rule adherence
  • task boundary awareness
  • completion criteria verification
  • safer autonomous execution controls

This feedback is based on a real multi-agent development workflow where Codex and Claude were used for independent review and verification.