CaSE: Conversations about Software Engineering Agent Harness, State of Play, Risk and AI Company Culture
Mar 17, 2026
Birgitta Böckeler, distinguished engineer at ThoughtWorks who writes on GenAI and AI-assisted delivery, digs into the current state of AI-assisted coding. She explores context engineering, agent teams and harness engineering. They cover risk thinking, small batches, cognitive load and how platform teams can enable safe, practical AI use in software work.
AI Snips
Chapters
Books
Transcript
Episode notes
Start Simple And Build Context Files Iteratively
- Start with plain agents, then iteratively add rules and skills instead of importing large shared rule sets upfront.
- Build rules from recurring mistakes or repetitive steps and share them with the team once proven useful.
Subagents Keep The Main Conversation Lean
- Sub-agents let an agent spin off dedicated contexts for research so the main session stays concise.
- Cloud Code spins explore sub-agents to read many files and then report summaries back to the main agent.
Gastown Showed Swarms Are Experimental And Costly
- Birgitta recounts that Steve Yegge's Gastown and Claude Flow spurred public swarm experiments but remain experimental and token-expensive.
- Gastown uses orchestration patterns like a mayor persona to schedule and aggregate agent outputs.
