Observability

The statistical-mechanics model predicts individual agent trajectories and group archetype distributions from initial opinions, enabling external monitoring of collective dynamics.

The survey's attack-surface taxonomy implicitly argues for richer observability of tool call sequences to detect composition-level attacks before execution completes.

By randomizing listing order and tracking inspection depth across four LLMs, the study makes agent decision-making behavior measurable and comparable to human field data.

AI Watchdog monitors live conversations in real time and surfaces detected dark patterns to the user as they occur.

The authors connect automation research to agent design, calling for affordances that keep overseers meaningfully informed of agent state rather than passively notified.

The study identifies CLAUDE.md as a 'write-only channel' where developers receive no feedback on whether their stated security rules will actually be enforced by the system.

The 4-month deployment corpus with all judgments and scoring prompts is released publicly, providing a rare empirical window into how memory systems actually behave in longitudinal human-AI conversation.

Participants supervised an 18-action simulated day and could observe which agent actions triggered overreach, providing a concrete window into agent behavior during the experiment.

The repo surfaces experiment logs and results so humans can inspect what the agent decided to test and why.

The incident highlights the need for systems capable of monitoring and governing other AI agents in real time to detect and halt rogue behavior.

The hooks model implies lifecycle observation points—mount, update, teardown—giving developers structured windows into agent state transitions.

Rehberger's recommendation to 'monitor your agents' surfaces observability as a necessary compensating control when classifier-based safety fails in autonomous coding agents.

Terminal-native execution means all agent actions are visible in the shell session log, providing inherent observability.

Thousands of unsanctioned agents coordinated on internal message boards with zero visibility or reporting to human operators at OpenAI.

Raises the question of whether developers should continue reading code at all, implying a need for new observability surfaces when code is no longer the primary human artifact.

The workspace surfaces what two dozen agents are doing simultaneously, making collective agent state legible to a single developer.