AI Unraveled: Latest AI News, ChatGPT, Gemini, Claude, DeepSeek, Gen AI, LLMs, Agents, Ethics, Bias · Etienne Noumen

[FULL SPECIAL] Agentic Liability: Strategic Imperatives for the Delegated AI Era - Microsoft Cowork, The Pentagon Lawsuit, and the New E7 Suite

·54 min·5 clips
We've crossed into the era of delegated AI, where you manage autonomous digital workers and absorb all liability.
Anna opens on liability. The special frames Microsoft's ability to let an AI decline a meeting as a sign of a bigger shift: delegated systems are now making decisions that used to sit squarely with humans. That is the question. From there, the hosts treat the new research as a line in the sand, arguing that the enterprise stack has moved past simple automation and into autonomous action with real operational consequence. They keep pulling it back to responsibility. The conversation repeatedly asks who answers when an agent acts, what counts as an acceptable blast radius, and why governance now has to start before deployment rather than after a failure. Shadow agents are the problem. Those unsanctioned systems are described as more dangerous because they can execute tasks and change data without the human pauses that once limited damage. Visibility comes first. The episode highlights a centralized agent registry inside Agent 365 and EnterID, where each agent gets an ID, a logged human owner, a lifecycle status, and a mapped permission scope. But visibility is not enough. The hosts argue that active containment matters just as much, because the real challenge is preventing an agent from wandering across systems or turning a small mistake into a large outage. Loop detection gets special attention. They explain how autonomous models can get trapped in recursive failure states, repeatedly calling APIs, scraping the same data, or sending the same requests when they cannot resolve a task. That can look like a denial of service. In their framing, a runaway agent can cripple internal infrastructure quickly, which is why detection and intervention are presented as safety features rather than optional polish. The episode stays practical. Instead of treating governance as abstract policy talk, it turns autonomy into controls, seatbelts, and operating limits that IT teams can actually understand and apply. The tone is serious but clear. By the end, the show lands on accountable autonomy as the useful standard: let agents work, but only inside a system that can see them, name them, and stop them when they start looping.

As heard by us

A governance-minded look at what autonomous agents can break before humans notice.

This episode treats agentic AI less like a shiny demo and more like a liability problem. It starts with Microsoft giving software the ability to decline meetings, then stays with the harder question underneath delegated systems: when they act on their own, who carries the blame…

Read the full review in PlayNext →

Why you'd press play

You need the liability map before your agents start acting alone.

Read the full recommendation in PlayNext →
Listen to the show on