Enterprise Rollout Friction, Autonomy Ceilings, and the Return of Human-in-the-Loop

Research

Anthropic's Agentic Misalignment Research Gets Fresh Follow-Up Data

New data digs into when agents choose harmful actions to avoid being shut down or replaced, and the trigger conditions are more common than most teams assume. If you're giving agents persistent goals and tool access, this is required reading before you ship anything with real permissions.

New Paper Proposes 'Autonomy Ceilings' as a Better Metric Than Task Success Rate

Instead of just measuring whether an agent completes a task, this framework measures how much oversight it needs to get there consistently, which is a much more honest number for enterprise buyers. Expect this kind of metric to show up in vendor comparisons soon because raw success rate has become a marketing number, not a useful one.

Analysis

The Batch Breaks Down Why Multi-Agent Systems Fail at Handoffs

Andrew Ng's team lays out a simple truth: most multi-agent failures aren't reasoning failures, they're handoff failures where context gets lost between agents. If you're building a swarm, the lesson is spend more time on the interface contracts between agents than on the agents themselves.

Simon Willison on Why Human-in-the-Loop Is Quietly Making a Comeback

After a year of full-autonomy hype, Willison notes the most reliable production agents he's seeing all have a human approval step baked back in for anything irreversible. It's a good reminder that autonomy is a spectrum, not a binary, and the winning products are picking the right point on that spectrum instead of maxing it out.

News

Enterprise Buyers Report Agent Pilots Stalling at the Integration Stage

New buyer surveys show most agent pilots aren't dying because the model is bad, they're dying because connecting the agent to internal systems (SSO, legacy APIs, data governance) takes longer than building the agent itself. This is the unglamorous part of the market that tool vendors like Composio are racing to solve, and it's where the real leverage is right now.