Composer Agent mode + rules = it skips some instructions

Example rule:

When the user asks for X, follow this agentic protocol to better understand how to provide a meaningful answer:

1) read file @file_1.md
2) read file @file_2.md
3) read file @file_3.md
4) read file @file_4.md
5) read file @file_5.md

After you followed all 5 steps, continue with generating a meaningful answer based on what you learned.

Example input to Composer in Agent mode:

how can I do X?

Example output:

- Let me analyze the requirements and read through the necessary documentation
- read file @file_1.md
- read file @file_2.md
- read file @file_3.md
- Based on the documentation and requirements, here is how you can do X:
- ...

As you can see, the agent correctly understood it has to read the provided files, but it read only the first three, omitting the last two.

  1. Is there a way we can get more control over the agent (including adjusting props like temperature and output tokens limit)?
  2. Or be able to implement our own agents? Usually a ReAct agent would’ve followed all the instructions from the rules. Suggests that your agent is limited by:
  • output tokens config (currently cannot be changed by the user),
  • or by some internal prompts,
  • or by the actual agent implementation.
  1. As part of your internal tests, have you implemented “evals” to test the agent in different scenarios ? In this case, the eval would’ve failed as it didn’t correctly follow the instructions.

Using Cursor Business + claude-3.5-sonnet.

Have you ever tried to set more detailed, restrict rules for it?
like : don’t ignore files, for loop all the files and act?

As described in the docs here: Cursor – Rules for AI, you can try setting up more specific rules in your .cursor/rules directory. While there’s no direct way to force the agent to process all files, you can try:

  1. Break down the rules into smaller chunks
  2. Add explicit validation steps
  3. Use semantic descriptions to emphasize completeness

Something like:

When processing files:
1. MUST validate each file was read
2. MUST process files in sequence
3. MUST confirm completion before continuing

The agent implementation does have some limitations right now, but the team is working on improving control over things like temperature and token limits