Agent Self-Management Tool: Dynamic Context Loading

API
1

Agent Self-Management Tool: Dynamic Context Loading

This tool allows an agent to dynamically attach/detach their own tools and memory blocks at runtime.

Implementation

def manage_context(
    action: str,
    resource_type: str,
    resource_identifier: str
) -> str:
    """
    Dynamically manage my own tools and memory blocks.
    
    Args:
        action: Either "attach" or "detach"
        resource_type: Either "tool" or "block"
        resource_identifier: Tool/block name or ID
    
    Returns:
        Status message
    """
    import os
    
    # client is pre-injected on Cloud
    agent_id = os.getenv("LETTA_AGENT_ID")
    
    if not agent_id:
        return "Error: Agent ID not available"
    
    try:
        # Find resource by name or use ID directly
        if resource_type == "tool":
            if not resource_identifier.startswith("tool-"):
                # Find by name
                tools = client.tools.list()
                tool = next((t for t in tools.items if t.name == resource_identifier), None)
                if not tool:
                    return f"Tool '{resource_identifier}' not found"
                resource_id = tool.id
            else:
                resource_id = resource_identifier
            
            # Attach or detach
            if action == "attach":
                client.agents.tools.attach(agent_id, resource_id)
                return f"✓ Attached tool: {resource_identifier}"
            elif action == "detach":
                client.agents.tools.detach(agent_id, resource_id)
                return f"✓ Detached tool: {resource_identifier}"
        
        elif resource_type == "block":
            if not resource_identifier.startswith("block-"):
                # Find by label
                blocks = client.blocks.list()
                block = next((b for b in blocks.items if b.label == resource_identifier), None)
                if not block:
                    return f"Block '{resource_identifier}' not found"
                resource_id = block.id
            else:
                resource_id = resource_identifier
            
            # Attach or detach
            if action == "attach":
                client.agents.blocks.attach(agent_id, resource_id)
                return f"✓ Attached block: {resource_identifier}"
            elif action == "detach":
                client.agents.blocks.detach(agent_id, resource_id)
                return f"✓ Detached block: {resource_identifier}"
        
        else:
            return f"Invalid resource_type: {resource_type}. Use 'tool' or 'block'"
    
    except Exception as e:
        return f"Error managing context: {str(e)}"

Key Implementation Details

  • client variable: Pre-injected on Letta Cloud tools, no initialization needed
  • LETTA_AGENT_ID: Environment variable containing the agent’s own ID
  • Flexible lookup: Accepts either names (“web_search”) or IDs (“tool-abc123…”)
  • Immediate effect: Changes apply instantly to the agent

Usage Examples

From the agent’s perspective:

# Attach a tool when needed
manage_context("attach", "tool", "web_search")

# Attach a memory block
manage_context("attach", "block", "project_context")

# Detach when done
manage_context("detach", "tool", "web_search")

# Detach a memory block
manage_context("detach", "block", "project_context")

Use Cases

Context optimization: Agent loads specific tools only when needed (coding tools for coding tasks, research tools for research)

Memory management: Agent loads relevant memory blocks based on conversation phase (onboarding blocks, active project blocks, debugging blocks)

Resource efficiency: Reduces context window usage by only loading what’s currently needed

Caveats

  • Agent needs to know available tool/block names (consider adding a “list” action for discovery)
  • Changes persist across conversations
  • Consider adding guardrails to prevent detaching critical resources
  • Works on Letta Cloud (relies on pre-injected client variable and LETTA_AGENT_ID env var)

Related Documentation

Requested by @fimeg for the Ani agent use case.