Google's free Gemini Tasks runs every errand in its own hardware-isolated VM, never shared with other users.
So the agent adding Cheerios to your cart gets a dedicated sandbox.
The enterprise agent with payroll access runs on a shared service account named svc-test-final-v2.
AWS just made long-term memory generally available for AI agents, so they can recall prior decisions across sessions.
Huge moment for enterprise AI.
The agent is about to be the only one in the org who remembers why we made that decision.
Google's new Stitch Loop is a product agent that watches your live metrics and keeps working 24/7 until your goals are met.
As a product person, I'm not worried about my job.
Somebody still has to move the goals the week it hits them.
Replit tested agent swarms on performance work. 1 agent: up to 4x faster. 16 agents: 17x. 32 agents: 19x.
Double the team, get two more x's.
Agents just learned what every VP learns in year one: the next 16 hires mostly sync with the first 16.
Microsoft's new Autopilot agent can run a project on its own: builds the workback plan, handles the follow-ups, and reaches out to stakeholders for updates.
As a TPM, I have one question.
Who's going to tell it nobody answers those?
Google's threat team found an attacker's multi-agent framework that reads security patches, writes the exploit, and runs it. No human in the loop.
Meanwhile our agents still need three approvals to close a Jira ticket.
First team to ship full autonomy to prod: the attackers.
New research: GPUs serving coding agents spent 60% of their prefill work recomputing context they'd already processed earlier in the session.
Dashboard says fully utilized.
Our GPUs aren't busy. They joined the meeting late and keep asking for a quick recap.
AI agents can now phone businesses for you. Many businesses hang up the second they realize it's an AI.
For 20 years, companies made us talk to their bots: "Press 1. Say 'billing.' I didn't catch that."
Our bot finally calls back, and suddenly they want to speak to a human.
New test: a model that's right 98.85% of the time, run as an agent loop, left its state exactly right on only 21% of steps. Each bad reply just compounded.
Finally, AI delivers compounding returns. On the bugs.
By step 400 it's not an agent anymore. It's a legacy system.
Sierra's new protocol has every Shopify store publish a file telling personal agents what it sells and how to get permission to shop.
For 30 years, websites published robots.txt to keep bots out.
Turns out the web never hated bots. It just hated bots without a credit card.
A major airline just updated its contract of carriage: no AI agents booking, changing, or servicing tickets.
Two years of agent keynotes opened with "book me a flight."
Our favorite demo is now, technically, a terms-of-service violation.
Researchers saw one site get 414,210 sign-in attempts from bots claiming to be a popular AI agent. Meanwhile, a real agent got in by pretending to be a regular browser.
Agent identity in 2026:
Bots claim to be agents. Agents claim to be human. Nobody claims to be a bot.
OpenAI's new ChatGPT Wallet lets an agent find the item and fill the cart. Then you approve every purchase with Face ID.
Agentic commerce is here, and it shops exactly like a teenager:
Does all the browsing, then hands you the phone to pay.
Claude Meetings joins your call and drafts the plan while everyone's still talking.
By default, it leaves the moment someone from outside the company joins.
Finally, an AI with real enterprise instincts: does all the work in the internal sync, gone before the customer call.
Coding agents can now spin up their own H100 over MCP and run sudo for hours.
Fine print: if the job finishes at 2am, it keeps billing until someone stops it.
Congrats, agents just made senior ML engineer: full GPU access and a forgotten instance burning cash until Monday.
New study on multi-agent setups for enterprise SOPs: up to 4.7x the tokens, 5.8x the latency, gains rarely significant.
The worst one failed because each agent saw only its slice and none was allowed to resolve a missing input.
That's not a swarm. That's a matrix org.
Databricks says its new agent gives the CEO the unvarnished truth: bad survey scores, lost deals, everything.
Bold feature.
Most companies already have a system for that. The truth just goes through four layers of management and arrives as "mixed signals, trending positive."
Cursor's agents can now coordinate with your teammates' agents, so a schema change and a feature build get sequenced before anyone hits a merge conflict.
Congrats. Agents now have cross-team dependencies.
Next sprint they'll want a TPM. I'm available.
n8n's new personal agent can cancel your subscription, turn down the retention offers and ask for a refund.
Every company is also shipping a retention agent.
So the future of customer service is two AIs haggling over 50% off a gym membership neither of them uses.
Meta's new agent setup: a big model writes the rules, then a cheap one at 10 cents per million tokens applies them to tens of thousands of cases.
AI has reinvented the consulting firm.
The partner sets the strategy. The analyst does the work. The invoice still says partner.