What's Shocking
- Stripe Minions ships 1,300+ PRs/week with zero human interaction during execution — Slack message in, merged PR out. Meanwhile OpenClaw requires a human to type three separate commands for one PR.
- Two-speed feedback saves entire CI cycles — local lint in <5 seconds catches type errors before a minutes-long CI run even starts. The agent never wastes a CI round on a missing import.
- Harness engineering inverts the improvement target — don't make the model smarter, make the environment more constrained until the model can't fail at predictable things. Every agent mistake becomes a permanent rule, linter, or deterministic node.
- Project management and memory are the same problem — OpenClaw's session-memory hook saves the last 15 messages as markdown after every task. Completed work becomes searchable context for future work.
See also: Memory Architecture (how agents store and retrieve knowledge) | Patterns That Work (blueprints, doom loops, harness engineering)
Theme: PM at scale reveals the audience problem. Stripe ships 1,300 PRs/week because its blueprint state machine removes the agent from predictable steps. OpenClaw requires three human commands for one PR because its 24-channel routing brain was built for audience reach, not workflow automation. The autonomy spectrum (fully autonomous → human-steered → agent-as-tool) maps directly to the multi-tenancy question: more autonomous agents need stronger isolation, and stronger isolation enables serving more untrusted users.