Agents & Automation
Agent architectures, orchestration, tool use, workflows, and operator patterns.
-
Agents & AutomationRead story
Can AI systems check each other? Researchers have proposals.
Researchers propose having AI systems that carry out tasks predict and check each other. A person could settle disputes. MIT Technology Review reports the idea, but risks include systems working together to fool the check. It has not been proved as a real-world safety control.
MIT Technology Review · Nov 3, 2026 -
Agents & AutomationRead story
Show the AI path it did not take
A study with 48 people tested a way to show paths an AI system did not choose during data analysis. Seeing an alternative helped people rely on the system less blindly. The test was not a school or workplace rollout.
arXiv · Nov 3, 2026 -
Agents & AutomationRead story
When AI acts for you, control points shape the failure
Researchers studied 695 public reports of failures by AI systems that can take actions. They found 12 ways people and systems can lose track of each other. How often a person can step in matters.
arXiv · Oct 30, 2026 -
Agents & AutomationRead story
An AI ran a mouse-brain injection. The stop button mattered.
In a reported test, a vet opened a mouse’s skull. A human approved the plan. Then AI chose where to inject a drug. A button could stop the machine.
WIRED · Oct 31, 2026 -
Agents & AutomationRead story
Lots of AI shopping use. Few confirmed orders so far.
Meta reports many people trying Muse for shopping. Merchants told WIRED they saw few orders from AI shoppers. Those are different measures, not a sales rate.
WIRED · Oct 31, 2026 -
Agents & AutomationRead story
Palantir’s CTO says AI gains belong to the whole job
In an interview, Palantir’s Shyam Sankar says firms should judge AI by whether a full job improves. This is a vendor’s view, not a test of results.
Stratechery · Oct 29, 2026 -
Agents & AutomationRead story
A personal AI tool reportedly reached a defense network
WIRED says a contractor ran a personal AI tool on a defense network for two months. Then it was found and patched. The public excerpt does not say what the tool could reach. It does not confirm that any data left the network. The reported access matters; a data breach is not confirmed.
WIRED · Oct 29, 2026 -
Agents & AutomationRead story
In one game, past choices beat what users said
An early research abstract describes a game with 42 users across 1,260 rounds. AI predictions based on conversation did not beat those based on past choices overall. The finding is about that game, not whether workers will use a new tool.
arXiv · Oct 27, 2026 -
Agents & AutomationRead story
A big AI math run still needs proof
Import AI relays a researcher’s account of AI-assisted math work: about $135,000 in estimated token costs and roughly 600,000 lines of Lean code. Neither figure was independently checked here.
Import AI · Oct 26, 2026 -
Agents & AutomationRead story
Make repeat work repeatable
Sabouri and colleagues made repeat steps in AI-assisted analytics runs reusable. Their set tests found more consistent tool calls. They did not test the method in a live office.
arXiv · Oct 24, 2026 -
Agents & AutomationRead story
Ask what finished, not how busy the AI looks
Mittal and colleagues map problems people face when they delegate work to AI tools. A tool may keep running without moving the task forward. Their research does not give a workplace failure rate.
arXiv · Oct 26, 2026 -
Agents & AutomationRead story
AI pitches news stories. Editors are not buying it yet.
WIRED reports that AI tools are sending story pitches to editors. One maker paused a tool after it wrote to public officials. The editors interviewed had not taken an AI pitch.
WIRED · Oct 26, 2026