TASIP-003: Human-in-the-loop steps

[ryw]

Status: Draft — open for input · Category: Big bet

Summary

Let an agent pause mid-run to ask a person for a decision or input through a structured form, then resume with the answer — unlocking workflows that need human judgment at a checkpoint (approval, disambiguation, sign-off) without abandoning automation.

Why

Plenty of useful work is "mostly automatable" with one human gate in the middle. Today you either fully automate (and accept the risk) or don't automate at all. A clean pause/resume with real forms is the missing middle.

Sketch

• An agent can emit a "need input" step with a schema (fields, choices, free-text).
• The run pauses; the request shows up for a human to act on.
• On submit, the answer is injected back and the run continues from where it paused.
  (Provisional — the shape changes based on the discussion.)

Open questions

• How long can a run stay paused, and what happens on timeout?
• Who can answer — anyone, the owner, a named approver?
• How does this interact with schedules and Slack-launched runs?

How to help

React 👍 if you want this, and comment with your use case, constraints, or pushback. This is a draft Tembo Agent Studio Improvement Proposal (TASIP); your input shapes it before it becomes scheduled work.