Problem and consequences
Pull requests don’t fit the shape of agentic development. A PR is a request with the wrong information. Not enough context – the why, the resonsing, how this came to be, alternate paths considered, rounds of review its been through, level of rigour. Missing the agentic information.
Shape of the solution
New primitives that capture the context, conversation, and commits of an agent session. Different information. Lighter weight and smaller than PRs. Continuous approval and merge. Better context and review tools.
Sessions are the unit of work. Vonversation with an agent that let to outputs: code commits, documents. We need to see the process. Like a scientific lab notebook. Still need sign-off, multiple eyes on work, audit log, historical context for the future.
Checkpoints - light weight review and approval gates
Trails - audit trails of exactly what happened, archived as historical context for future work
Chronicles – natural language summaries, explainers, and contextualised overviews of an agent session. Made to orient yourself and others.
LLMs are great information compressors and expanders.
Design challenges
- LLM are unreliable narrators problem
- How does async and sync work work?
- Who constructs the chronicle and when
- What are the boundaries of a session - when does it start and end
- Session forking
- cross-session review