The Audit Problem: Who Watches the Watcher-Agents?

Research

Decentralized Task Allocation Still Breaks Under Load

New analysis shows agent fleets coordinating via market-style bidding start thrashing once task volume spikes past a threshold nobody bothered to stress-test before. If you're running a zero-human shop on autopilot, this is the kind of paper you read twice — the failure mode is silent until it isn't.

Tools

AutoGPT Ships Better Failure Logging, Not More Autonomy

Latest release leans hard into observability — structured logs, retry diffs, task lineage — instead of flashier autonomy claims. Smart move: the bottleneck for running agents unsupervised was never capability, it was knowing what broke and why after the fact.

Task Expiry and Dead-Task Cleanup Finally Standard

Buried in the changelog: automatic expiry for stalled tasks so zombie processes stop eating compute and API budget silently. Small fix, huge deal for anyone who's woken up to a surprise cloud bill from an agent fleet that never got the memo to stop.

News

YC Keeps Funding the 'No Headcount' Pitch

Another batch of YC-backed startups pitching agent-run ops teams instead of hiring — support, ops, even basic BD now framed as agent fleets with a human on top as 'orchestrator.' The thesis is getting less contrarian and more default, which means the arbitrage window is closing fast.

Analysis

Herding Bias Resurfaces as the Quiet Killer of Agent Fleets

Follow-up work digs into how agents copying each other's decisions (herding) compounds errors faster in zero-human setups where there's no human in the loop to notice the drift. It's basically groupthink for bots, and it's way harder to catch without a dashboard built specifically to flag correlated mistakes.

Self-Correcting Agents Need Self-Correcting Audits Too

The same paper's later sections argue that self-improving agent loops need an independent audit agent — not just better prompts — or you end up with a fleet that's confidently wrong in a consistent direction. This is the uncomfortable part of 'zero human' nobody wants to admit: you still need a human-designed checkpoint somewhere.