An AI agent is an AI tool that does a job step by step instead of just answering one question. I opened every source below myself before writing this. When a number or claim comes from the company selling the product, I say so.
Google announces Agent Compute Engine, a home for AI agents
Google Cloud announced Agent Compute Engine on 22 October at the AGNTCon + MCPCon event. In plain words, it is a service that gives each AI agent its own small, sealed-off computer (a microVM) to work in. Google's blog says it "will be available in preview in the coming months."
Google says each microVM has its own guest kernel, its own identity and short-lived credentials. Service passwords are added outside the microVM, so the agent itself never holds them. A related open-source project, Agent Substrate, is available now for Kubernetes on GKE. Google claims it gives "as much as 10x higher compute density" than traditional compute.
The benchmark in the figure comes from Google's blog: 200,000 agents, 500 activations per second, and suspend and resume in under 500 milliseconds. Google also describes an Agent Gateway that it calls "air traffic controller for client-to-agent and agents-to-anywhere traffic," plus a registry of approved agents and tools.
My take: Google sells this, so every number here is Google's own, and I found only Google's blog post as a source. The idea is sensible: if an agent can act on its own, it should work in a sealed room and never hold the keys. But there is no price yet and no independent test.
AWS open-sources Infinity, a way to run tens of thousands of agents
Amazon Web Services published Infinity on 23 October. It is a research runtime, which means the engine that keeps agents running. It is released under the Apache 2.0 open-source licence as part of Hydro, a project between AWS and UC Berkeley. The AWS post calls it "an early research release" shared to gather feedback.
The trick is simple. When an agent is waiting for a tool to answer, Infinity pauses it so it uses almost no memory. When the answer arrives, it wakes the agent up with its saved state.
In AWS's own benchmark, 80,000 agents made 3.2 million completions on a single thread of an AMD EPYC processor. Memory grew by about 100 KB per idle agent, so roughly 75,000 agents fit in 8 GB. The test used a simulated model, with each agent running 20 turns, so it says nothing about real model speed or cost.
My take: AWS made this and ran the benchmark itself, so treat the numbers as AWS's claims. It is early research, not a product you can buy. Still, it shows the big cloud companies now think agents need their own kind of engine.
TechCrunch lists 24 AI agents you can text
TechCrunch published a guide on 24 October, written by Lauren Forristal, to 24 AI agents that live in your text messages. You message them like a person, and they remember context, connect to your apps and do tasks like booking, scheduling, emailing and shopping.
Prices vary a lot. TechCrunch lists Ask Proxy at $1 a day for personal use, Folk Pro at $8.33 a month and Martin from $21 a month. Others, like Instinct, have no price listed. Several are still beta, testing or invite-only. The most prominent, Instinct, has a $10 billion valuation after a $1 billion round.
The privacy notes matter most. TechCrunch says Instinct's growing independence has raised privacy and security concerns. Wajo gives its agent its own email, phone number and payment card so you do not share your own details. Underdog keeps its model and data on your own device and is invite-only for now.
My take: This is one outlet's roundup, so I did not test any of these agents. If you try one, start with something low-risk and ask what it can touch before you connect your email or payment details.
Block shows Dessa, a voice model that can listen while it talks
Block, the company led by Jack Dorsey, announced Dessa from its AI research lab on 23 October, according to Voice AI Space. TechCrunch covered it on 24 October. Dessa is a speech model that listens and talks at the same time, and Block says it can laugh, pause and sound tired.
Block says it trained Dessa entirely on human speech, using a single rack of NVIDIA GPUs, and that it can run locally on a MacBook with an M4 Max chip or better. It is not publicly available. Voice AI Space reports a waitlist and early access for a small group of safety researchers. TechCrunch says Block has invited researchers to test whether a voice agent could deceive people.
Block argues that most voice AI turns speech into text, which loses emotion. It also suggests an on-device model could lower costs for its business and consumer products.
My take: Almost every claim here comes from Block. Both reports say there are no independent benchmarks, and the demos were picked by Block. It is a research preview, not a tool you can use today.
What I take from this week
- Agents are getting their own infrastructure. Google's sealed-off microVMs and AWS's pause-and-resume runtime are both about running many agents safely and cheaply.
- Agents are moving into everyday apps. A 24-item list of texting agents shows how fast this is spreading to normal users, with privacy questions to match.
- Almost every number this week is from the seller. Google's, AWS's and Block's figures are claims, not independent proof.
If you use AI agents, the practical question is the same each week: what can it touch, who set the limits, and who checked the claims. I will keep reading the sources myself and say plainly when a number comes from the seller.
Sources
- Google Cloud Blog: Introducing Agent Compute Engine and expanding agent governance
- AWS Open Source Blog: Introducing Infinity, a runtime for extreme agent concurrency
- TechCrunch: Here are the top AI agents that can live in your text messages
- TechCrunch: Block's new voice model sounds eerily human
- Voice AI Space: Block launches Dessa, a full-duplex speech model trained only on audio