Working Surfacelink · 23 Oct 2026
02
Three small wind-up carts roll along a narrow track towards a row of level-crossing barriers, and the last barrier stays down until a figure seen from behind lifts it by hand, the whole scene drawn with wide empty margins on every side.
Thoughtworks

How Sensible Harness can help teams create an AI-first delivery environment

Vitória Regina Ragazzon Toebe, Amy Pellegrini and María José Lalama · 22 October 2026

Company story · agents · critique & review · Human approval

Read the original
Takeaway

A Thoughtworks team built Sensible Harness to run software delivery with three AI agents whose limits the tool enforces, and a person approves each plan before code is written.

Summary

A Thoughtworks team built Sensible Harness, a command-line tool, at the company's 2026 AI hackathon to prepare code repositories for delivery work with AI agents. Vitória Toebe, María José Lalama and Amy Pellegrini describe AI agents for business analysis, development and testing, each held to steps it cannot skip. Thoughtworks now uses the tool for its own delivery work.

Key points
  • The development AI agent drafts a plan that quotes each acceptance criterion, and a person approves it before any code exists; the approved plan is posted on the issue.
  • The testing AI agent marks any scenario it cannot run as blocked, never as passed, and hands it back to a person.
  • Each skill is pinned to an AI model suited to its difficulty, and a cost report is posted to the ticket after every task.
  • A code audit grades the repository from A to F on ten dimensions, including test quality, security and AI governance.
Implication

Engineering teams can treat the instructions and skills their AI agents use as code that is versioned, tested and reviewed. This is Thoughtworks' account of a tool it built and promotes.

Plays
Engineering

Post a cost report to the ticket after an AI agent finishes each task, so the team can see where the spending goes.

Engineering

Have a person approve the AI agent's implementation plan before any code is written, and post the approved plan on the issue for the next person or agent.

Derived by Working Surface from the article; more in the Playbook. Source line: We’ve learned that standards hold when they’re enforced as gates the agent can’t skip, not offered as guidelines it might follow.

Source issue

23 October 2026: teams get dependable work from AI agents by writing the rules first and keeping the checks with people and ordinary code that the agents cannot override.