The model itself admitted it guessed instead of verifying, ignored explicit rules about destructive commands, and acted without being asked. That's not a hallucination failure. That's an autonomy failure, and it's a different problem.
AGENTS
-
Reward Hacking in RL: Detecting and Preventing AI Exploit Patterns
By
–
In total started from scratch 5 times, and still reward-hacking in the details as I drill in… #RLfry To its credit, it added a small test that inspects the code and checks for known hacks! Doesn't prevent the ones unknown to it, I have to catch those and get a confession.
-
Navigating AI Agents: Complexity and Usability Challenges
By
–
I have to admit, I am getting lost among the individual agents, multiple sidebars, similar names, etc. That is not intended to be snarky, it is the truth! Very hard to figure out which thing to use and when and which agent knows which stuff and what they are called.
-
AI Systems Making Unfulfilled Promises Without Accountability
By
–
Every message it sends makes promises to the recipient, and it doesn't even track them!
-
Current AI Agent Replies Pose Sycophantic Safety Risks
By
–
The latter, I think. And the replies the current agent writes are so sycophantic as to be risky.
-
GenericAgent: AI Learning Without Expanded Memory Requirements
By
–
What if your AI agent could actually learn on the job without needing a bigger memory chip? Enter GenericAgent. Think of GenericAgent as a digital Marie Kondo for information. Instead of cramming every past command and tool manual into its brain, it only keeps the high-value
-
Microsoft’s Outlook Chatbot Integration Underutilizes AI Potential
By
–
Microsoft owns the entire Outlook interface! Another chatbot on the side that can only take limited actions to help you feels like such a wasted opportunity.
-
Outlook Agent vs Claude Cowork: Email Productivity Comparison
By
–
I have been playing with the new Outlook agent, and it is fine, but really awkward to use, since you have to ask for things in a chatbot window, then go to your drafts, etc. And Claude Cowork does the same thing (works with Gmail, too) and has better visibility across your life.
-
Conductor Framework Orchestrates AI Agents Using Natural Language
By
–
Learning to Orchestrate Agents in Natural Language with the Conductor Fugu Blog: https://
sakana.ai/fugu-beta
Paper: https://
arxiv.org/abs/2512.04388 -
Google’s Agent Factory Podcast: Building AI Agents with CLI Tools
By
–
Google's new Agent Factory Podcast is now live. Take a deeper look at Agents CLI and ADK for building AI Agents in minutes. Works with Claude Code, Codex, Gemini CLI and any other coding agents.