How to make it do research or simple search?

by StanislavKalishenko - opened Jul 26

Jul 26

I've checked different variants of this LLM (Q4_K_L, Q4_K_M) and it's not clear how to force it to do a search. It generates broken links, hallucinates and does not search anything. Possibly it's needed to sent it instructions in proper way because I did it as simple message. Like "Do research about Claude Code's subagents".

alandao

Menlo Research org Jul 26

Hi please retry with the Q8 version to see if you have the same issue

StanislavKalishenko

Jul 26

Didn't help. Here is the answer on "Do a research about Claude Code's subagents mode. How to set up them and run?".

"
Okay, the user is asking about Claude Code's subagents mode and how to set it up and run. Let me start by recalling what I know about Claude Code.

Claude Code is part of the Anthropic Claude series, specifically designed for code generation and execution. Subagents in this context probably refer to agent-based systems or multiple agents working together within a single interface.

First, I need to confirm if there's an official documentation on subagents mode for Claude Code. Since I can't browse the internet, I'll have to rely on existing knowledge.

Subagents might allow users to create multiple small agents that handle different tasks within the same code environment. Setting them up could involve defining each agent with specific roles and then integrating them into a larger system.

To set up subagents, steps might include creating definitions for each agent, configuring their interactions, and ensuring they have access to necessary resources or APIs. Running them would likely require invoking each agent's functions through the Claude Code interface.

I should also consider potential challenges, like managing multiple agents, ensuring they don't conflict, and handling communication between them. Maybe mention using specific parameters or commands when initiating subagent execution.

Claude Code (part of Anthropic's Claude series) is a powerful tool for code generation and execution within an AI environment. While it doesn't have a direct "subagents" mode like some other frameworks (e.g., Kubernetes or Docker), there are conceptual approaches to simulate or implement subagents-like behavior using its capabilities. Here's a structured breakdown of how subagent mode might be approached, setup, and execution in Claude Code:

🔍 What is Subagent Mode?

Subagent mode refers to the ability to break down complex tasks into smaller, independent agents (e.g., for code generation, data processing, or system integration). Each subagent can execute specific functions or roles within the larger coding environment.

In Claude Code:

Subagents could be defined as modular components (e.g., a "data cleaner," "API caller," or "code optimizer").
They operate within the same codebase but handle distinct tasks.
Communication between subagents is managed via APIs, shared data structures, or predefined workflows.

🧠 How to Set Up Subagents in Claude Code

1. Define Agent Roles

Create a definition file (e.g., subagents.yaml) that specifies each agent's role and capabilities:

subagents:
  - name: "DataCleaner"
    description: "Processes raw data for cleaning and normalization."
    functions:
      - function: clean_data
        parameters:
          input: str
          output: dict

2. Integrate with Claude Code

Use Claude Code's API or CLI to register subagents in the environment:

claudemy agent add --name DataCleaner --description "Data cleaning tool" --function clean_data

3. Configure Agent Interactions

Define workflows between agents (e.g., a "CodeOptimizer" that uses a "DataCleaner"):

workflows:
  - name: "Pipeline"
    steps:
      - agent: DataCleaner
        input: raw_data
      - agent: CodeOptimizer
        input: cleaned_data

4. Access Subagents via API or SDK

If using the Claude Code SDK, you can invoke subagent functions programmatically:

from claudemy.sdk import Agent

cleaner = Agent("DataCleaner")
cleaned_data = cleaner.invoke({"input": "raw_input_string"})

🚀 How to Run Subagents

1. Start the Environment

Ensure the Claude Code environment is running:

claudemy run --mode subagent

2. Execute Workflows

Use commands or scripts to trigger subagent workflows:

claudemy workflow execute --workflow Pipeline

3. Monitor and Debug

Check logs or status of agents:

claudemy agent status DataCleaner

💡 Tips for Effective Subagent Mode

Modularize Code: Keep each subagent focused on a single task.
Use Version Control: Track changes in subagent definitions.
Error Handling: Implement retries or fallbacks for failed subagents.
Security: Restrict access to sensitive subagents (e.g., API keys).

🛠️ Potential Limitations

No Built-in Subagent Mode: Claude Code’s interface may lack native subagent support, requiring manual scripting or external tools (e.g., Python scripts).
Resource Constraints: Multiple agents could strain computational resources.
Communication Overhead: Efficiently managing inter-agent communication is critical.

📚 References

For detailed documentation, refer to:

If you have specific use cases (e.g., integrating with external APIs or handling parallel tasks), let me tailor the setup further!"

As you see, it made up everything (links are not working, there is no such claudemy, etc.). And in thinking there is "start by recalling" that shows that it event didn't try to search.

Note: I'm checking it in LM Studio.