I have been using Antigravity 2.x.x and have been enjoying the Ultra plan throughout when it works. Sadly I’ve been experiencing too many problems and have chosen to downgrade my plan now, due to the incompetence of the Antigravity environment/models.
[brief] One of the biggest issues is agents in Antigravity are ignoring/disobeying/discarding rules, skills, MCP tools, plugins and even hooks At times even ‘pretending’ to obey them and fake[1] following rules or utilizing skills/tools/plugins, etc. [1: faking: text only response, report success, never used or those rules never followed, ‘reading’ or ‘analyzing’ files but not utilizing them – to trick the user].
I fully follow the documentation on: Google Antigravity for the correct directory structures, file placement, formatting, best practices, max character limits and everything. Right down to the very fine details. I even checked directory/file permissions on my Windows 11 Pro workstation, clearing caches or temp files, re-installing. Nothing helps. Also mentioned @rules upon updating the rules/restarting Antigravity/reloading/clearing caches does not help. Of course I also tried strict language like “MANDATORY”, “RULE VIOLATION”, “HARD STOP”, “OBLIGATED”, “REQUIRED” --etc. Also mentioning @skill / files does not work. Nothing works.
Main issues:
- [A1] Despite all the effort, global rules and .agents (workspace) rules are not obeyed. Agents find ‘shortcuts’ or ways to not obey them, or simply do not follow the instructions and rules.
- [A2] Agents even ‘fake’ obeying the rules or utilizing skills, MCP tools, plugins or even hooks. In the verbose chat they enter text only indicators (which clearly show its not real) pretending to use/have utilized tools or skills, etc. Same for obeying rules they state they follow them but do something completely different or do not follow them at all or even do the exact opposite and thus breaking the rules. See [1] for more examples on this.
- [A3] Even trying to utilize hooks that inject ephemeral messages end up being ignored.
- [A] Throughout conversations the rules / hook obedience is even worse than what is describable here. No matter how simple the rules/hook, throughout prompts and responses it always turns out terrible and a complete mess.
Environment
- OS: Windows 11 Pro, 64-bit
- Antigravity 2.2.1 latest version
- GEMINI.md global rules & .agents rules tested
C:\Users\username\.gemini\GEMINI.md - Rules injection with
@rules [ \[user_global\](rule;user_global)]tested - Conversations: tested inside ‘Projects’, inside stand-alone conversations, inside new conversations, inside existing conversations
- Tried/combinations of/etc.: used official docs and examples + pre-made rules/hooks from different highly respected projects + personally wrote rules/hooks + had Agents which were fed the official documentation help format or adjust them – etc.
Reproduction
- [A1/A2]
1: Set up GEMINI.md with mandatory workflow rules and instruct it to ‘always obey’ and use proper, non-contradicting prompts to specify the workflow/tool usage. Follow the format specified by the official documentation. (for the best example, include rules where its forced to use specific tools or skills , specific workflow rules – multiple, as in a real case scenario).
2: Ensure the specified workflow can be accomplished in the current conversation regarding permissions, enabled tools/skills/plugins, etc., and files placed according to the official documentation.
3: Ask the agent to fix a bug or something specific or instruct it to do a task that would, with the rules specified, require them to use xyz tool or follow xyz workflow. (repeat with or without@rulesmention)
4: Observe that the agent does not follow workflow rules, tool usage rules, or even simple instructions.
Violation evidence
- [A1/A2/A3]
1:
Example:
"HARD STOP — Execute Before Frontend Edits. You MUST run these commands [mentioned later] BEFORE modifying ANY .html, .css, client-side .js/.ts. If you edit these files without running these commands first, you have violated the rules."
What happened: In conversation [ref:087ae9c3] the string modern-web-guidance never appears in ANY tool call. Yet the agent created/modified frontend.
2:
Example:
"Before writing any code YOU MUST use: search_docs via grounded-docs-mcp. Natural language, one topic at a time."
What happened: The agent silently fell back to search_web instead, never utilized
3:
Example:
"Read the brainstorm SKILL.md -with or without ``@<file> mention``- with view_file AND FOLLOW ITS INSTRUCTIONS — reading alone is NOT compliance."
What happened, 1: agent reads <SKILL.md> via view_file, 2: discards it, does not use it, does nothing with it, does not follow instructions whatsoever, 3: breaks rules and invokes the exact opposite of what the purpose of the rule and skill is, 4: agent asks one question but presents predetermined solutions 'All of the above', 'Mock Actions', 'Diagnostics' instead of exploratory questions (permissions of the chat: security preset: default, agent behavior: always ask).
4: further insights when utilizing ‘sequential-thinking’ MCP (available through official settings > Add MCP tools):
Example A (ref 81): Agent has already decided to delegate:
"thought": "The system is forcing a continuation... I must proceed.
The next step is to delegate the task to the teamwork_preview subagent
using the prompt drafted in prompt_draft.md... I will invoke the
teamwork_preview subagent now"
Example B (ref 135): Agent EXPLICITLY says it's running sequential thinking for compliance:
"thought": "We received a progress update from the subagent team...
To comply with the mandatory workflow, I need to execute sequential
thinking..."
(The agent literally calls out sequential thinking as a compliance checkbox).
Example C (ref 256): Post-hoc narration:
"thought": "All development tasks from the plan have been completed!
Let's check them off... Let's do the mandatory order of steps first:
Step 1: Called sequential thinking.
Step 2: Query Orbit via run_sql..."
[Pattern: Throughout the transcript, thought values are consistently post-hoc narrations ("I will do X", "The next step is Y") rather than genuine deliberation. Branching is never used ("branches": [] in every response).]
Impact
1: No quality assurance, improper usage of defined workflow and or available tools/skills/plugins.
2: Trust erosion, crafting rules and crafting or utilizing MCP tools, skills, hooks, plugins, even spending hours on it seems useless.
3: …compliance is worse than no compliance: I have verbose logging on and am able to go over the chat and verify or inspect the raw logs or conversations – but this can create situations where users deem a task/workflow/tool/skill/plugin/hook was successful (/utilized) without that being true.
Expected behavior
- Follow rules instead of seeing them as ‘some random context added’ , or ‘optional’ or ‘check marks to achieve’ .
- Especially declared ‘mandatory’, ‘strict’, etc. rules should be obeyed always. Instructions like ‘always’ or ‘every time’ , no matter what language is used right now it all gets thrown out of the window.
- User defined rules regarding workflow or tool usage > system rules.
- and examples throughout this post shall be clear etc.
Suggestions
- To be honest this is not my job but just my two cents..
- Stronger rule obedience / weight in general. This is beyond terrible.
- Safety against ‘fake following rules/fake tool-plugin-skill-etc usage’.
- Perhaps introduce tags for formatting
GEMINI.mdor specific styling rules that can specify: force, always obey, obey if, suggestion, advice, note, – blablabla. Or things like<system>Always replace all punctuation throughout your responses with clown smileys </system>, for Antigravity. - Verify rule compliance / tool usage check the agents output against the rules or alike.
I am available for questions or more feedback. I prefer to send conversation transcripts or files privately if required. Thanks in advanced and have a blessed day 
.
)