Questions and Debates.
Themes I'm following and experimenting with. If you're in NYC thinking about similar stuff, let's connect.
Agents
Workflows or big model?
DAG-based workflows and durable execution, or frontier model + loop + generic tools + broad latitude?
Invest in local apps and open weight models or just wrap APIs?
Open weight models lag by a few steps, but does the absolute gap narrow over time? Privacy and cost benefits vs. batch cloud compute and frontier capabilities. Bet on local or wrap APIs?
What are the best context engineering idioms?
RAG or grep? When to prune, cultivate, compact? Use a generic harness like Claude Code and focus on domain knowledge?
SDLC and DX
Sync, async, or supervised async?
Once you can containerize and parallelize coding agents, what actually works? Supervise 2-3 sessions? Fire up async work and review PRs? We are the bottleneck.
How do I not be the bottleneck?
How to set up kanbans, triage issues, and decompose work for async and semi-sync agent workflows.