Parallel Tool Calling

When an agent calls multiple tools, Letta can execute them concurrently instead of sequentially.

Parallel tool calling has two configuration levels:

• Agent LLM config: Controls whether the LLM can request multiple tool calls at once
• Individual tool settings: Controls whether requested tools actually execute in parallel or sequentially

Model Support

Parallel tool calling is supported for OpenAI and Anthropic models.

Enabling Parallel Tool Calling

Agent Configuration

Set parallel_tool_calls: true in the agent’s LLM config:

1
const agent = await client.agents.create({
2
  llm_config: {
3
    model: "anthropic/claude-sonnet-4-20250514",
4
    parallel_tool_calls: true
5
  }
6
});

Tool Configuration

Individual tools must opt-in to parallel execution:

1
await client.tools.update(toolId, {
2
  enable_parallel_execution: true
3
});

By default, tools execute sequentially (enable_parallel_execution=False).

ADE Configuration

Agent Toggle

1. Open Settings → LLM Config
2. Enable “Parallel tool calls”

Tool Toggle

1. Open the Tools panel
2. Click a tool to open it
3. Go to the Settings tab
4. Enable “Enable parallel execution”

Execution Behavior

When the agent calls multiple tools:

• Sequential tools execute one-by-one
• Parallel-enabled tools execute concurrently
• Mixed: sequential tools complete first, then parallel tools execute together

Example:

Agent calls:
  - search_web (parallel: true)
  - search_database (parallel: true)
  - send_message (parallel: false)
Execution:
  1. send_message executes
  2. search_web AND search_database execute concurrently

Limitations

• Parallel execution is automatically disabled when tool rules are configured
• Only enable for tools safe to run concurrently (e.g., read-only operations)
• Tools that modify shared state should remain sequential