Why Long-Running Tasks Deserve First-Class Treatment in MCP

If you’ve been following the Model Context Protocol (MCP) — the open standard that lets AI agents and tools talk to each other — you’ve probably heard whispers about a new proposal called SEP-1391.

So what is it, in plain English?

Think of SEP-1391 as a “patience upgrade” for MCP.

Until now, MCP assumed that when an agent calls a tool, the job finishes instantly — like asking your calculator for 2+2.

But in the real world, AI doesn’t just answer questions; it runs tasks that take time — like generating reports, training models, or waiting for human approval.

SEP-1391 introduces a simple idea:

You can start a job now, get a tracking ID, and check back later when it’s done.

That small change makes MCP dramatically more useful for real enterprise workloads — especially in industries like advertising, marketing, and data where asynchronous jobs are the norm.

What SEP-1391 Adds

In short: it gives MCP a way to handle long-running, stateful jobs gracefully.

• Start a job → Get a token back immediately
• Track progress with tools/async/status
• Fetch results later via tools/async/result

It’s like getting a FedEx tracking number for your AI task — you know where it stands without waiting on the line.

New concepts include:

• Async mode: invocationMode: "async"
• Job states: submitted, working, input_required, completed, failed, canceled
• Polling hints: how often to check and how long results are stored (keepAlive)

Why It Matters (Especially for AdTech + MarTech)

Almost every serious marketing workflow runs long:

• Data ingestion & clean-room joins — hours of processing
• Lift or incrementality studies — days or weeks of data collection
• Creative QA & variant generation — human approvals mid-flow
• Budget pacing or optimization — algorithms iterating in the background
• Compliance or audit bundles — async reports that must stay retrievable

These are not instant operations. They need checkpoints, human input, and proof of completion.

That’s exactly what SEP-1391 brings to MCP.

The New Lifecycle

Every async task can now move through these states:

submitted → working → input_required → completed → failed → canceled

That middle one — input_required — is especially powerful.

It lets a human pause, approve, or modify a job without breaking it.

The Bigger Shift

This evolution turns MCP from a simple “call-and-response” system into a true orchestration layer — one that understands time, state, and human checkpoints.

For teams building agent-based automation in AdTech, MarTech, or enterprise AI, that means:

• Jobs can span hours or days
• Humans can review or intervene mid-run
• Results stay auditable and retrievable

In other words, MCP just learned how to operate in the real world.

In One Line

If SEP-0001 made MCP interoperable,

SEP-1391 makes it durable.

Agents can now wait, resume, and prove they’ve done the work.