The rapid proliferation of autonomous AI agents makes governing agentic systems urgent. We introduce the competence circuit, denoted $IKC + KC + C$, which decomposes an agent's real‑world impact into three pathways:
- Intelligence‑mediated discovery (IKC)
- Procedural know‑how (KC)
- Direct control (C)
Using the canonical fast‑takeoff debate as a stress test, we formalize the sufficient conditions for intelligence‑mediated fast takeoff: when fast catastrophic supremacy occurs, the following three sub‑conditions must hold simultaneously:
- Intelligence is high enough to select hypotheses efficiently (i.e., epistemic intelligence $E_{hyp}=\frac{\text{effective hypotheses}}{\text{total hypotheses}}$)
- Knowledge acquisition rate $K_{rate}$ is sufficiently fast
- Direct control over critical capabilities can be exercised with low latency $C_{latency}$
Most scenarios rely not on raw intelligence alone but on rapid knowledge acquisition or control attainment. We operationalize epistemic intelligence as hypothesis‑selection efficiency and validate the framework via stochastic simulations on NK fitness landscapes, covering six scenarios that span different combinations of the three pathways.
The simulations support an analytic bound: $$ S{intelligence}\le \frac{I{max}}{I_{competitor}} $$ where $S_{intelligence}$ is the intelligence‑driven speed‑up multiplier, $I_{max}$ the highest intelligence level, and $I_{competitor}$ the competing intelligence level. This bound holds regardless of task difficulty.
Fast‑takeoff dilemma: a large multiplier emerges only when the hypothesis space is noisy, yet the absolute timescale remains limited by the time $T_{test}$ required for capability‑critical tests, so a high multiplier does not necessarily imply rapid progress.
Institutional chokepoints become disproportionately effective when mandatory authorization delays $\tau_{auth}$ dominate the negligible time agents spend on unproductive search. Consequently, we argue that restricting control mechanisms and enforcing knowledge boundaries are more potent safety levers than merely limiting intelligence, a principle that applies across all capability levels. Purely digital domains weaken these bottlenecks, strengthening the case for architectures that intentionally "put physics in the loop."
Review: The paper combines theory and simulation to shift AI safety focus from "how smart" to "how much control and knowledge flow"—a pragmatic guide for policymakers.