The Changelog: Software Development, Open Source Opus 4.5 changed everything (Interview)
148 snips
Feb 27, 2026 Burke Holland, an engineer on GitHub Copilot who experiments with AI coding workflows, talks Opus 4.5, Copilot, and agentic development. He recounts one-shot wins, building small apps with models, and how agents fit into real developer workflows. They debate tool costs, why humans still matter for shipping, and how agents enable rapid prototyping and more builders.
AI Snips
Chapters
Books
Transcript
Episode notes
Set Checkpoints And Human Approvals For Long-Running Agent Tasks
- Automate what you can but keep a human-in-the-loop for ambiguous decisions; set up notifications and checkpoints (e.g., Telegram) for agent progress.
- Burke aims for agents to act like a team lead, messaging when they need direction and continuing after approval.
Classify Tasks And Route Agents Based On Complexity
- Tailor agent workflows by task difficulty: classify tasks as easy/medium/hard and route simple queries to quick answers while reserving planning and subagents for hard tasks.
- Burke's Anvil agent does this classification and delegates to different models and subagents accordingly.
Spec To Ship Overnight Invoicing App Ledger
- Adam described using Claude (via Augie) to spec a Rails invoicing app overnight and waking to a working V1 called Ledger with PDFs, emails, and login.
- He combined a detailed spec, Claude prompts, and Augie automation to produce a usable app by morning.
