All posts

Engineering

8 min

How we decide when an agent should stop and ask

Confidence scores aren’t enough. Here’s the three-part check every Envoy agent runs before it acts on its own.

Sam Whitfield

Head of Applied AI

worm's-eye view photography of concrete building

The cost of a wrong answer

Most agent failures aren’t spectacular. They’re a refund to the wrong card or a polite email to the wrong person. So we built the decision to pause around the cost of being wrong, not just the model’s confidence.

Three questions before every action

  • Is this action reversible within the tool?

  • Is it inside every limit the workspace has set?

  • Would a reasonable teammate ask first?

If any answer is no, the agent pauses and sends a question with the context needed to decide in one tap.

Keep reading

a blurry orange and yellow background with a white border

Product

Introducing Envoy 2.0: agents that hand off to each other

Agents can now pass work to each other mid-run, with the same approvals and receipts — so one brief can span support, finance and ops.

a close up of a piece of white paper

Guides

The approval limit playbook

A practical guide to setting guardrails your finance and security teams will actually sign off on.

a blurry image of a car from a window

Company

Receipts, not dashboards

Why every Envoy run ends in a short report instead of another chart nobody opens.

Create a free website with Framer, the website builder loved by startups, designers and agencies.