writing.
Monitors that watch an AI agent's session miss harm that fires after it ends
Monitors that score an AI agent's commands as it runs them caught none of four attacks that hide the harm in a scheduled job or system trigger. Listing and resolving those triggers at session end caught all four, with no false alarms on four honest sessions.
ai safetyai controlmonitoringOctober 14, 20268 min readWhat Broke When I Ran Thirteen Claude Code Sessions on One Always-On Machine
Long-running agent sessions failed where they acted on a state they could not see: an account balance, an approval, another session's progress, or their own compacted context. Here is each failure and the change that fixed it.
agentsworkflowclaude codeOctober 14, 20266 min readSportsbook Spillovers: Kalshi's Forecasting Markets Tilted Toward Longshots After the Sports Launch
After Kalshi launched sports markets in January 2025, taker order flow in its non-sports markets tilted toward longshots and longshot returns deteriorated. We document the shift with 593 million trades, multiple-testing corrections, and complete robustness reporting.
researchprediction marketseconomics+1August 5, 202618 min read