Software Engineering Radio - the podcast for professional software developers SE Radio 730: Birgitta Boeckeler on Harness Engineering for AI Agents
18 snips
Jul 22, 2026 
Birgitta Boeckeler, a Distinguished Engineer at ThoughtWorks who specializes in AI-assisted software delivery, discusses harness engineering for coding agents. She explains guides and sensors, how they integrate with CI/CD and linters, and practical harness elements like observability and evals. Conversations cover coordinating multiple agents, building trust in agent outputs, and the evolving responsibilities of engineering teams.
AI Snips
Chapters
Transcript
Episode notes
Try The Agent Vanilla Before Extending It
- Start using coding agents plain‑vanilla before adding custom harness layers to learn what the base tool already does.
- Birgitta recommends iterating: experience default behavior first, then add targeted guides and skills step by step.
Guides And Sensors Are The Harness Core
- Birgitta splits harness content into guides (feed‑forward instructions like Agents.md) and sensors (feedback tools like static analysis or tests).
- Guides set norms and context; sensors let agents self‑correct using LLM or computational checks before human review.
Turn Linters Into Actionable Agent Sensors
- Use existing tools (ESLint, SonarQube, Semgrep) as sensors and customize rule messages so agents can make informed judgment calls.
- Example: overwrite max‑lines messages to ask the agent to consider refactoring or justify exceptions.