The More I Think About Autonomous Agents, the More I See Latency as a Security Problem
The biggest risk for autonomous agents may not be making the wrong decision. It may be failing to finish the right decision before the market moves.
Newton’s design walks through how an Intent is evaluated before settlement, how operators independently execute the policy, and how their decisions become a BLS-aggregated attestation.
What I think is easy to miss is why the policy language matters so much.
As more protocols experiment with autonomous agents and machine-speed execution, the question isn't just whether policies are correct—it is whether every authorization decision can be completed before the opportunity to act disappears.
Newton avoids bringing that problem into authorization by refusing to let arbitrary programs define permission logic in the first place. Policies are written in Rego—a declarative language where rules are evaluated against defined inputs such as "input", "data.params", and "data.wasm".
That predictability allows independent operators to evaluate and sign the same authorization result within a bounded decision process.
That changes how I think about "pre-settlement." Pre-settlement isn't simply a faster compliance pipeline. It follows from choosing a policy model where every operator can deterministically evaluate and sign the same result before liquidity moves.
The interesting trade-off isn't speed. It's expressiveness. Some authorization logic becomes harder—or impossible—to encode once predictable evaluation boundaries are enforced by design.
In autonomous systems, a correct decision that arrives too late can become an incorrect outcome.
The design depends on every policy remaining deterministic enough for independent operators to reach the same decision before the execution window closes.
Bounded evaluation limits policy complexity. The question is how much expressiveness a protocol should sacrifice to guarantee decisions finish before liquidity moves.