Examples / 02 — Connectors

02 — Connectors

Real agents need tools. The connectors helper turns common expert types into the raw experts shape the API expects — fully typed. This example builds a two-agent system: a focused symptom-extractor sub-agent wired into an orchestrating triage agent.

Run it

# Minimum — uses sub-agent connector only
npm run connectors

# With the web-search registry expert
USE_WEB_SEARCH=1 npm run connectors

# With your own MCP server
MCP_URL=https://my-mcp.example.com MCP_TOKEN=sk-... npm run connectors

Connector types

All connectors are created with the connectors helper exported from @corti/agent-sdk. Pass the result array to agents.create({ connectors: [...] }) or agent.update({ connectors: [...] }).

Walkthrough

1 · Create the sub-agent

The symptom-extractor is a focused, narrow-purpose agent. Its system prompt is written to ensure deterministic output — always a comma-separated list, never a question, never prose.

const symptomExtractor = await agents.create({
  name: "symptom-extractor",
  description: "Extracts chief complaint and key symptoms from a clinical note.",
  systemPrompt:
    "You are a symptom extractor. Given a clinical description, respond with " +
    "ONLY a comma-separated list of the patient's chief complaint and key symptoms " +
    "(e.g. `severe headache, photophobia, neck stiffness`). Never ask for clarification.",
});

This agent has no connectors — it relies only on its training. The narrower the system prompt, the more predictably it behaves when called as a tool.

2 · Wire it as a connector

connectors.fromAgent({ agentId: symptomExtractor.id })

This produces a cortiAgent expert binding. When the orchestrator decides to use this tool, the platform routes the call to symptomExtractor and returns its response. The orchestrator never sees the sub-agent's raw API calls — it only sees the reply text.

3 · Build the orchestrator

const orchestrator = await agents.create({
  name: "triage-orchestrator",
  description: "Triages a clinical note using the symptom extractor.",
  systemPrompt:
    "You are a clinical triage assistant. Pass the full clinical note to the " +
    "`symptom-extractor` connector to get a symptom list, then write a brief " +
    "one-paragraph triage recommendation.",
  connectors: [
    connectors.fromAgent({ agentId: symptomExtractor.id }),
    // Opt-in extras:
    ...(process.env.USE_WEB_SEARCH === "1"
      ? [connectors.registry({ name: "web-search-expert" })]
      : []),
    ...(process.env.MCP_URL
      ? [connectors.mcp({ mcpUrl: process.env.MCP_URL })]
      : []),
  ],
});

The system prompt references the connector by its description — the orchestrator knows it should delegate symptom extraction. The platform presents available connectors to the orchestrator as tools it can call.

4 · Run with a raised timeout

const reply = await orchestrator.run(
  "62-year-old presents with sudden severe headache, photophobia, and neck stiffness.",
  { timeoutInSeconds: 180 },
);

The default timeout is 60 seconds — sufficient for a single-agent call. Orchestrated calls fan out to sub-agents internally, which can take longer. Passing timeoutInSeconds: 180 gives the orchestrator enough headroom.

Note that agent.run() is a one-shot convenience wrapper. The same options are available on ctx.sendMessage() and ctx.sendText().

Full code

Source: examples/ts/02-connectors.ts

/**
 * 02 — Connectors.
 *
 * Attach MCP servers, registry experts, and other Corti agents to an agent
 * via the typed `connectors` helpers.
 *
 * By default this example wires a sub-agent (`connectors.fromAgent`) in —
 * always available, no external services. Set `USE_WEB_SEARCH=1` to also
 * attach the registry `web-search-expert`, or `MCP_URL` to attach your own
 * MCP server.
 *
 * Run: `npm run connectors`
 */
import { AgentsClient, connectors } from "@corti/agent-sdk";
import { makeClient } from "./_client";

async function main() {
  const agents = new AgentsClient(makeClient());

  // A small sub-agent we will wire into the main agent as a "cortiAgent" connector.
  const symptomExtractor = await agents.create({
    name: "symptom-extractor",
    description: "Extracts chief complaint and key symptoms from a clinical note.",
    systemPrompt:
      "You are a symptom extractor. Given a clinical description, respond with ONLY a comma-separated list of the patient's chief complaint and key symptoms (e.g. `severe headache, photophobia, neck stiffness`). Never ask for clarification. Never add prose. If the input is sparse, extract whatever symptoms are mentioned.",
  });

  // The orchestrator composes connectors. The sub-agent is the default path;
  // the registry expert and MCP server are opt-in.
  const orchestrator = await agents.create({
    name: "triage-orchestrator",
    description: "Triages a clinical note using the symptom extractor.",
    systemPrompt:
      "You are a clinical triage assistant. Pass the full clinical note to the `symptom-extractor` connector to get a symptom list, then write a brief one-paragraph triage recommendation (urgency, likely differentials, immediate actions). Never ask the user for clarification — work with whatever is provided.",
    connectors: [
      connectors.fromAgent({ agentId: symptomExtractor.id }),
      ...(process.env.USE_WEB_SEARCH === "1"
        ? [connectors.registry({ name: "web-search-expert" })]
        : []),
      ...(process.env.MCP_URL
        ? [
            connectors.mcp({
              mcpUrl: process.env.MCP_URL,
              ...(process.env.MCP_TOKEN ? { token: process.env.MCP_TOKEN } : {}),
            }),
          ]
        : []),
    ],
  });

  // Orchestrated calls fan out to sub-agents / experts, so raise the per-call
  // timeout beyond the SDK's 60s default.
  const reply = await orchestrator.run(
    "62-year-old presents with sudden severe headache, photophobia, and neck stiffness for the past hour.",
    { timeoutInSeconds: 180 },
  );
  console.log("Status:    ", reply.status);
  console.log("Reply:     ", reply.text);
  console.log("Artifacts: ", reply.artifacts.length);
}

main().catch((err) => {
  console.error(err);
  process.exit(1);
});

What to expect

Next steps