Effective AI-Native Development with Harness Engineering

Most AI coding demos are greenfield theater. Your real work isn't. It's a 12-year-old codebase with patchy tests, a tangle of frameworks from three eras, and conventions nobody can fully explain. Independent research shows AI tools can actually slow experienced engineers on exactly this kind of code, unless the harness around the model is built deliberately.

This six-hour, hands-on tutorial teaches harness engineering: the discipline of designing everything around a coding agent that isn't the model itself. You'll work in a pre-instrumented legacy sandbox, adding real features with Claude Code while a live dashboard tracks coverage, complexity, and code health on every commit. The goal isn't faster output. It's confident, repeatable change you can defend with metrics.

Across three arcs and six tasks, you'll bootstrap a project harness with instructions, skills, and curated tools, add features behind a characterization-test safety net, and externalize the harness so your team gets the benefit, not just one engineer. We use Claude Code as the anchor. The principles transfer to Cursor, Copilot, and any agent that follows.

You'll leave with a one-page Monday-morning plan for your own brownfield repo, working examples of every harness component, and an honest framework for measuring whether AI is actually making your codebase better.

For engineers, tech leads, and architects already using coding agents and ready to take them seriously in production code.

Key Takeaways

1 Build the strategic context of AI-native Development at the enterprise level and articulate a case for Harness Engineering

2 Educate on key concepts and principles of Harness Engineering

3 Practice hands-on harness engineering skills across 5 iterations, including reflections - 4.5 hours

Speaker

Zichuan Xiong

Head of AIOps @Thoughtworks

Zichuan Xiong has been a Principal at Thoughtworks since 2008, currently leading the AI practice within Thoughtworks Managed Services, where he architects agentic AI solutions for software operations. He partners with Chief AI, Data, and Technology Officers to translate emerging AI capabilities into measurable business impact across technology and product organizations. His core expertise spans AI, Domain-Driven Design, Data Mesh, and organizational design and transformation.

Read more

Speaker

Premanand Chandrasekaran

Head of Technology @Thoughtworks

Premanand (Prem) Chandrasekaran is a technology leader and change agent with a solid track record of leading large technology teams and helping businesses deliver mission-critical problems while exhibiting high internal and external quality. In the past two decades, he has had the pleasure of helping a variety of clients and domains ranging from financial services, online retailers, education, and healthcare startups among others. His specialties include technical innovation, architecture, continuous delivery, agile/iterative transformation and employee development. When not fiddling with his trusty laptop, he spends time with his son ripping beyblades, playing video games and analyzing the nuances of cricket.

Read more
Find Premanand Chandrasekaran at:

Date

Friday Nov 20 / 09:00AM PST ( 7 hours )

Share

Share Share

Prerequisites

Attendees should have:

  • At least 3 years of professional software engineering experience in any language
  • Reading-level familiarity with Java (the sandbox codebase is Java/Spring Boot). You do not need to be a Java expert. The workshop teaches principles that transfer to any language, with explicit cross-language reflection at each step.
  • Some prior exposure to a coding agent such as Claude Code, Cursor, GitHub Copilot, or Cline. Beginners welcome; we start from first principles.
  • Comfort with Git basics (clone, branch, commit, push) and the command line

No prior experience required with harness engineering, MCP, agent skills, hooks, or subagents.

Hardware and setup (must be completed before the event):

  • Laptop with a current version of Chrome, Edge, Firefox, or Safari
  • Stable Wi-Fi connectivity (the workshop runs entirely in a browser-based cloud development environment)
  • GitHub account with Codespaces enabled. Free tier is sufficient. We will provide the repository link in pre-work.
  • Active Claude Code access via Anthropic. **Claude Max ($100/month) or Anthropic API access is strongly recommended.** Claude Pro may hit usage limits during a 6-hour intensive session.
  • Completion of the pre-workshop verification script. We will email this to registrants 10 days before the event. Running the script and submitting its output is required to participate.

What we provide:

  • A pre-instrumented legacy sandbox repository (forked from a real open-source project), one branch per attendee
  • A pre-configured GitHub Codespace with all workshop tooling installed
  • A live metrics dashboard showing test coverage, complexity, and code health
  • All slides, reference cards, and a Monday-morning checklist for applying the workshop in attendees' own codebases

What we do not provide:

  • Anthropic credentials or Claude Code subscriptions
  • Local IDE setup or laptop troubleshooting
  • Cellular hotspots or backup connectivity

A pre-workshop reading takes about 20 minutes and grounds the shared vocabulary for the day:

  • Birgitta Böckeler, *Harness engineering for coding agent users* (martinfowler.com, April 2026)

Optional but encouraged:

  • -One sentence submitted via the pre-work form describing the most painful AI coding friction the attendee has hit in a real brownfield codebase. We read all submissions and reference common themes during the opening.

On the day, attendees should bring:

  • A notebook or note-taking app for personal reflection. Each task ends with a structured reflection, and externalizing notes during the day improves transfer to attendees' own codebases on Monday.
  • Willingness to pair with someone they did not arrive with. We assign pairs to mix seniority and stack background, which is part of the design.
Save your place

Register