Skip to main content

Tool Registry

ToolRegistry.executeParallel() runs all tool calls from a single LLM turn in parallel and enforces a shared output budget.

executeParallel​

When the LLM requests multiple tools in one turn (e.g., read three files simultaneously), they all execute concurrently. The before_tool_call hook fires for each before any execute — rejected tools are excluded from the batch.

const results = await toolRegistry.executeParallel(toolCalls, {
  hooks,
  resultBudgetChars: 80_000,
})

Result budget​

The default budget is 80,000 characters total across all concurrent tool calls in a single turn. This budget is split evenly across the tools being executed:

3 tools executing → 80,000 / 3 = ~26,666 chars per tool

If a tool's result exceeds its per-call budget, it's trimmed and appended with:

[truncated — 45,231 chars total]

This prevents runaway tool results from consuming the entire LLM context window.

Per-tool maxResultChars​

Tools can declare a tighter limit than their share of the budget:

export const readFileTool: Tool<typeof schema> = {
  name: 'read_file',
  maxResultChars: 20_000,
  // ...
}

The actual per-call budget is:

Math.min(perCallBudget, tool.maxResultChars ?? perCallBudget)

A read_file tool with maxResultChars: 20_000 won't produce more than 20,000 chars even if the split budget allows more.

Adjusting the budget​

Pass resultBudgetChars in AgentLoopConfig:

const loop = new AgentLoop({
  // ...
  resultBudgetChars: 120_000,  // increase for tools with large outputs
})

isAvailable​

Tools that implement isAvailable() are checked at registration time. Tools returning false are excluded from the LLM's tool list — the model never sees them or tries to call them.

export const browserTool: Tool<typeof schema> = {
  name: 'browser',
  isAvailable() {
    return Boolean(process.env.PLAYWRIGHT_HEADLESS)
  },
  // ...
}