Progress from a Vercel AI SDK agent, and a board to watch it on

Run several steps and log each one

A step is one model generation plus the local tools it asked for. By default generateText and streamText stop after one step (isStepCount(1)), so set stopWhen to let the agent keep going. Other stop conditions are hasToolCall(...toolNames) and isLoopFinished().

import { generateText, isStepCount } from "ai";

const { text, steps } = await generateText({
  model,
  tools,
  stopWhen: isStepCount(10),
  prompt: "Upgrade the payments service to the new API",
  onStepEnd({ stepNumber, toolCalls, toolResults, finishReason }) {
    console.log(`step ${stepNumber + 1}: ${toolCalls.length} tool calls, ${toolResults.length} results (${finishReason})`);
  },
});

const allToolCalls = steps.flatMap((step) => step.toolCalls);

prepareStep runs before each step and can change the model, toolChoice, activeTools or messages for that step, for example to force a planning tool on step 0.

Seeing a run from outside the process

onStepEnd only runs in the process that started the agent. A run in a queue worker, a serverless function or a cron job has nobody reading its logs as it goes, and no way to ask you something and wait for the answer. It needs to report somewhere you can see. The same holds for agents built on the OpenAI Agents SDK or the Claude Agent SDK.

Connect Taskhold's MCP server

createMCPClient from @ai-sdk/mcp connects to remote servers over HTTP, which the docs recommend for production. A server-side agent can't do a browser sign-in, so it uses an API key: start the 7-day trial, open Settings > API keys and make one for this agent.

import { createMCPClient } from "@ai-sdk/mcp";
import { generateText, isStepCount } from "ai";

const taskhold = await createMCPClient({
  transport: {
    type: "http",
    url: "https://taskhold.com/api/mcp",
    headers: { Authorization: `Bearer ${process.env.TASKHOLD_API_KEY}` },
  },
});

try {
  const { text } = await generateText({
    model,
    tools: { ...(await taskhold.tools()), ...myTools },
    stopWhen: isStepCount(40),
    system: WORKFLOW,
    prompt: "Upgrade the payments service to the new API",
  });
} finally {
  await taskhold.close();
}

• Each run shows as a task on the board, with a bar split by step, the agent's label, how long it has run and what it's doing now.
• When the agent blocks with a question, it waits in taskhold_updates (up to 50 s per call) while you answer from any browser or your phone. Leave room in stopWhen for those waits. More on that in answering a blocked agent.
• tools({ schemas }) loads only the tools you list, if you want the agent to see fewer of Taskhold's.
• redirect defaults to error, so use the exact URL https://taskhold.com/api/mcp: the www host redirects.