Steve Yegge on fences, not sandboxes

A concrete, hands-on argument that large agent workforces need governance systems that capture institutional knowledge as explicit law, not just narrower sandboxes.

Logged at IST: 2026-08-24 23:01 IST

What it is: Steve Yegge's essay on what his Wheelhouse agent organization for Wyvern has taught him about governing fleets of AI workers.

Gist: Yegge's claim is that once Fable-class models get cheap enough for every company to run hundreds or thousands of AI employees, the control problem stops being mostly about tight sandboxes and narrow task wrappers. In his account, his agents gradually built something closer to a legal system: offices, roles, jurisdictions, rulings, case law, runbooks, gates, tripwires, authority envelopes, and mechanical checks that decide whether a change is legal.

The useful part is the organizational framing. Agents are amnesiac, replaceable workers coordinating through text, so the system needs durable written law: who may act, what precedent says, which rules have been ratified, and which fences politely refuse actions that lack the right authority or timing. Yegge argues that this is how tribal knowledge becomes machine-operable, and that every serious agent-heavy organization will need its own locally grown version rather than a generic transplant.

Newsletter angle: Strong follow-up to the agentic-engineering/org-design thread: governance for AI employees may look less like one perfect sandbox and more like institutional law, lightweight fences, and continuous maintenance of the rules agents use to coordinate.

{{}}