Writing
-
The Attacker's Verifier: Offense-Defense Balance Under a Shared Vulnerability Finder
A model that finds vulnerabilities below human cost looks like a gift to attackers. Model security as a race and the opposite falls out: a shared finder is a coin flip, patching compounds while exploiting depletes, and gated release is the defender buying a head start.
-
Swarm Economies: Market-Based Coordination for Autonomous Agent Fleets
When thousands of agents contend for tools, budget, and compute, central planners buckle. Internal markets - prices, auctions, and budgets - recast coordination as an equilibrium problem with provable welfare guarantees.
-
The Verifier's Advantage: Asymmetric Verification as the Engine of Reliable Autonomy
Reliability at scale will not come from better generators but from cheap, composable verifiers. We formalize the generation-verification gap and derive when stacking checks dominates scaling the policy.
-
Context as Compute: A Memory Hierarchy for Long-Horizon Agents
Treat the context window as the top of a managed memory hierarchy, not an infinite scratchpad. Eviction, prefetch, and a working-set model turn attention from a cost center into a scheduled resource.
-
The Orchestration Substrate: Toward an Operating System for Autonomous Agents
Agent frameworks are reinventing the operating system one ad-hoc loop at a time. A principled substrate - scheduling, a memory hierarchy, and capability-based isolation - is the precondition for fleets that run for months, not minutes.