Edgar Degas, Paul Lafond and Alphonse Cherfils Examining a Painting, c. 1878–80. The Cleveland Museum of Art, 1958.25. CC0.
Merge judgment is the bottleneck.
Good morning.
This week’s bookmarks flooded the desk with plausible agent work — then argued the same spine. Factories ship. Humans still merge. The bottleneck is morning triage and merge judgment, not another mega-agent brand.
Merge judgment is the bottleneck. Everything else is costume.
Factories ship. Humans merge.
Jean-Michel Lemieux calls quarterly roadmaps the fax machine of the agent era. Replacement: Continuous Compilation — keep a list of ideas agents compile overnight into working product. Write a short prompt. Let a base model run. Morning triage: keep, trash, or learn. Usually a few rounds until shape appears. When it works, you get months of old planning cadence in a week — and ideas you couldn’t have scheduled.
Same bottleneck on YAGNI with Nate Berkopec: infinite plausible pull requests haven’t made teams much faster, because deciding what can merge is still the hard part. One study found developers felt faster and measured slower. Start factories on changelog-class work. Track human touches and revert rate. Zach Lloyd splits the job: product taste that won’t fully delegate, and factory engineering that raises automation without pretending taste is free.
Matt Swanson · Zach Lloyd · Ben Holmes
Harness control plane — not model brand.
Claude Code mods are in-process TypeScript plugins that can rewrite the agent’s behavior, UI, and tools — deeper than skills or MCP add-ons. The AGENTS.md loader itself is a mod. Pi 1.0 ships a minimal MIT harness with Codemode, deferred tools, cache warming, and experimental durable sessions outside the terminal. OpenClaw Enterprise is an open control plane for persistent agents in sensitive environments — trust boundaries, audit, swappable harness and sandbox. Heinrich packages knowledge work as typed agent repos: objects, skills, and interfaces with IDE-like diagnostics, reusing coding harnesses instead of inventing a new one per domain.
ClaudeDevs · Pi · OpenClaw · Heinrich
Compete on harness and governance. Leave the model logo on the shelf.
Instruction hygiene + QA. A marketing toy.
Kun Chen: don’t dump the full aerospace Simplified Technical English ruleset into AGENTS.md. Sample about ten recent transcripts, score which clarity rules would have cut confusion, and write that subset into the instruction file. Treat AGENTS.md as living harness code optimized from traces. e2e ships open agentic testing — mix deterministic locators with agent steps, cache the green ones, bring your own model, web and mobile.
Secondary: HeyGen Video 1.0 as a cheap text/image-to-video API for demo and onboarding clips. Useful toy. Not the stack.
Try this
Three moves. Own the merge.
- Replace one roadmap slice with overnight compile and morning keep/trash/learn. Log human touches and reverts for a week.
- Ship one small in-process harness mod — a policy guard or instruction loader — before another prompt pack.
- Sample ten recent agent transcripts; promote only the clarity rules that would have cut confusion into AGENTS.md. Trial an agentic e2e smoke with action-cache replay.
Merge judgment is the bottleneck. Keep the triage. Leave the costume on the shelf.