capability-budgetBound what the agent may do, not just what it may not
An explicit allow/ask/deny policy plus rate, cost and time ceilings, enforced by the harness rather than requested of the model.
Four dimensions govern every action: scope (which files, tools, accounts), rate (writes or calls per interval), reversibility (irreversible actions need a human), and visibility (who is told, what evidence is kept). A capability budget encodes all four before the first unattended run, and the harness enforces them — the model cannot restrict itself, and a permission is a harness property, not a model property. Generalises the narrower irreversible-action gate to include ceilings on cost, rate and duration, which are the bounds that stop a runaway loop rather than a single bad call.
Requires: Classifying tools and targets by reversibility up front, and a runtime that actually mediates every call — a budget the agent can bypass is a suggestion.
Contexts: Autonomous agent, Coding agent, Customer support
Does it work?
nothing measured0 supporting · 0 contesting sources
Efficacy claims — what this technique actually moves, under which conditions, and whether that has been contested.
No efficacy claim filed yet. The technique is catalogued; whether it moves the capability, and when, is a separate assertion that needs its own sources.
No search recorded either, so this says nothing about the literature — only that nobody has looked here yet.
Code
No repository linked yet. Contribute one.