Tools

Tools are the action layer of Agent extensions, available during a Turn. Use built-in Tool groups when the Agent needs platform-supplied planning, Memory, or durable file operations instead of instructions alone.

How Tools are enabled

The Agent-level tools array contains Tool group IDs and defaults to []. At Turn resolution, each selected group expands to its declared Tool names; built-in Tools are off unless their group is enabled.

Selection does not provision a backing capability. file_operation_sandbox requires the Agent to have a tenant-owned Sandbox attached, while the Memory group uses the Agent-owned Memory store. Selected Skills and attached MCP connections add their own Tools separately.

Enable a Tool group

This backend example enables planning, asks the Agent to use it, consumes the Turn, and reads the resulting Session transcript.

import { BlazingAgents } from "@blazing-agents/sdk";

const client = new BlazingAgents({ apiKey: process.env.BLAZING_AGENTS_API_KEY! });
const agent = await client.agents.create({
  name: "Release planner",
  tools: ["write_todos"],
});
const turn = await client.chat({
  agentId: agent.id,
  message: {
    id: crypto.randomUUID(),
    role: "user",
    parts: [{ type: "text", text: "Use write_todos to plan a release." }],
  },
});
for await (const _chunk of turn.toUIMessageStream()) { /* consume the Turn */ }
const transcript = await client.sessions.messages(agent.id, await turn.sessionId);
console.log(transcript.data.at(-1)?.parts);

The final assistant message's parts include the write_todos Tool call and result before the Agent's response.

Built-in Tool groups

Skills are not another public Tool group. Selecting a Skill adds activate_skill and a constrained read path for supported Skill assets. Attached MCP connections contribute discovered remote Tools rather than entries in tools.

Tool activity and approvals

For a successful Session Turn, assistant UIMessage parts retain Tool calls and their results in the transcript. A sensitive call can instead pause for a Tool approval; approval applies to that exact call, not to its Tool group.

A failed or canceled Turn is still metered, but only a successful Turn changes the Session transcript.

Production considerations

• Enable only groups the Agent needs. Tool selection is capability exposure, not access control.
• Attach a Sandbox before selecting file operations; the API rejects an invalid configuration and execution checks it again.
• Expect Tool errors, Turn cancellation, and deadlines. Do not treat a proposed call as a completed side effect.
• Keep credentials, tokens, and other sensitive values out of Tool arguments, prompts, and application logs.

Related guides

• Configure file operations
• Create and attach a Skill

Reference

• TypeScript SDK: agents.create() and agents.update()
• REST API: POST /v1/agents and PUT /v1/agents/:agentId