Subcategory of Artificial Intelligence & Future of Work
Autonomous & Agentic Workflows
Evaluate when to delegate tasks to autonomous agents versus keeping them manual or human-led.
One Approval, Two Payments: The 24-Hour Trap
One paused approval answered by two workers can fire two payments.
Your Agent's Next Step: 11-14x Likelier to Fail
One wrong step does not stay one mistake. In a study of 2,518 agent runs, a single wrong step made the next step roughly 11 to 14 times likelier to fail.
Your Agent Fabric's 52.2% Blind Spot
Two measurements landed a day apart: one left 13 capability-steps reachable after their last justified use, the other lifted disclosure to an agent the…
Retry Budgets Fail: Agents Retry 60 of 80 Anyway
Uber cut the worst retry storm its network can produce from 25 hops to three by taking the decision to retry away from the caller.
Agent Timeouts: Pick 0.1%, Not a Round Number
Amazon sets a call timeout from an accepted false-timeout rate of 0.1%, then reads the downstream p99.9. Most agent stacks pick one round number instead.
270 Attacks: Your Green Tests Are Blind
Two research groups, a day apart, ran 6,368 coding-agent runs and 270 attack cases and found the same blind spot.
Long Context Is Not Memory: Why AI Agents Need Tiered State
When Chroma engineers tested 18 large language models (LLM) across eight lengths, scores dropped as the prompts grew — even on simple tasks.
Agent Memory Poisoning: Guard the Write, Not the Prompt
One simple text file is enough to take over a coding agent's memory.
The Retry Is the Bug: An Agent's Costliest Failure Mode
A corrupted value sent a ledger agent into 15,000 retries and a $50,000 bill. A 25,930-episode study shows the same retry duplicates payments while the agent…
The Tool-List Budget: How Many Tools One Agent Can Hold
Giving one AI agent a massive tool list crashes its accuracy to roughly one pick in seven.
Durable Execution Is Not a Transaction: The Missing State in Agent Tool Contracts
Across 98,291 tools built for the Model Context Protocol (MCP), an AI agent rarely knows if it can undo a mistake.
Your Multi-Agent Failure Has a Shape: Read the Topology Before You Swap the Model
When an agent stack fails, the common reflex is to buy a bigger model. Two new papers point to a different fix: the shape of the system.
The Agent Safety Gate Moves to Verified State
An \"approved\" AI action no longer means a safe result. Say an agent changes a bill to pay it.
Planner-as-Router: Fold the Model Choice Into the Plan
On Sept. 26, three researchers shared a new system called Planner-as-Router (PaR) that cuts AI agent costs by 44 percent.
Stop Piling Memory Onto AI Agents: Optimize Skills
Drop one hard check from an Artificial Intelligence (AI) agent's learning loop, and its test score crashes from 17.2 to 6.6.
MCP Skills Extension: Standardizing Agent Workflows
Skills — portable bundles of workflow instructions — became an official MCP extension on September 13, standardizing how agents discover, retrieve, and verify…