Fences, not Sandboxes
- Source
- yegge.ai
- Author
- Steve Yegge
- Date
Context
Steve Yegge argues that coordinating agents requires turning an organization's unwritten expectations into explicit rules. In the essay's account of Wheelhouse, the system used to build the game Wyvern, decisions and incident reviews become recorded rulings, and rules that keep being violated tighten from custom to warnings to written law to mechanical enforcement: programs that refuse an action or alert an operator. This is the author's own description of a system, not an independent . Essay.
The essay's term for those enforcement mechanisms is a fence: anything that turns an agent away when it is not supposed to be there, such as Wheelhouse allowing only one model to contact humans, enforced at the Slack and email boundaries. The argument concerns governing cooperative agents, and the essay explicitly says a fence is not a and will not stop a malicious superintelligence. The design proposal is to connect written policy to enforceable actions, with a lifecycle that proposes, ratifies, measures, amends and retires rules as the organization changes. Governance mechanism.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
- guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
Checking sign-in…
Loading comments…