noticias radar ia agentes de ia automatizacion robotica

Tuesday's radar was about who charges for AI; Wednesday's is about who already has it working. Toyota has shared internal details on how it brought over fifty agents to production and reduced the deployment of each from six months to four days—which is exactly the leap almost no one makes. Meanwhile, money is moving into the physical world—XPeng's robotics subsidiary raised over $900 million at a $6.3 billion valuation—programming with agents became 82% cheaper within Kiro, and serious discussions are beginning about who is liable when an agent spends your money.

Toyota has 50 agents in production and has dropped from six months to four days per agent

Toyota North America published yesterday the breakdown of its ToyotaGPT platform: more than fifty agents in production spread across manufacturing, supply chain, financial services, dealerships, and R&D, managed by a team of about 35 people. The "before and after" numbers are what matter: putting a new agent into production went from six months to four days, the effort dropped from six engineers to one per agent, and a line diagnosis agent compressed incident resolution from five or six hours to two or three minutes. The team works with an explicit threshold—each project must return between six and seven figures annually—and with inherited permissions: the agent only sees the data the user could already see. For your company: this is the counterexample to what we discussed last Friday, when almost everyone wanted agents and nine out of ten stayed in the pilot phase. What sets Toyota apart is not the model, it's that they stopped building agents one by one: a library of reusable skills, a single model gateway with alternative providers, and permissions tied to source systems rather than reinvented per agent. Translate this to your scale: if your second agent is costing you the same as the first, you don't have an AI problem, you have a platform problem, and the sign that you're doing well isn't how many agents you have, but how long the next one takes. And copy the threshold: if a case cannot justify an annual savings figure before starting, it's not the case to start with. Source

XPeng raises 900 million for its humanoid: AI money is moving to the physical world

XPeng's robotics subsidiary closed a round yesterday of more than $900 million at a post-money valuation exceeding $6.3 billion, the largest private "physical AI" deal in China to date. Led by IDG Capital, with participation from Gaorong Ventures and strategic investors Tencent and Alibaba; about 600 million comes from external investors, about 200 from a subsidiary of the group itself, and about 100 from its management team. The target is the IRON robot—76 degrees of freedom in the body and 21 in each hand—with mass production scheduled for the end of this year, initial deployment in the brand's stores and campuses, and commercial deliveries in 2027. XPeng stock fell 6.8% following the announcement. For your company: the headline is the robot, but the useful takeaway is where the capital is going. Until now, your automation bill depended almost entirely on the price of the token; from now on, a growing portion of industry money is financing hardware with factory schedules, and that means that throughout 2027 you will start receiving physical automation offers—warehouse, inventory, inspection—with the same rhetoric that software reaches you today. You don't need to do anything today except two things: do not accept a robot demo as proof of anything without a measured pilot on your own floor, and organize the digital side first, because a robot in a warehouse whose inventory is not properly recorded only moves an error faster. Source

Programming with agents costs 82% less inside Kiro: internal software ceases to be the bottleneck

OpenAI and AWS today introduced the GPT-5.6 family—Sol, Terra, and Luna—into Kiro, Amazon's development environment that converts a product specification into an implementation plan before writing code. In joint tests on Terminal-Bench 2.1, Terra completed tasks with a cost reduction of around 82% compared to the previous approach. Kiro maintains two controls that are the interesting part for a company: manual review points where someone accepts, adjusts, or rejects changes before they enter, and verification by properties—checking that the code meets defined rules, not just passing fixed tests. This comes on top of the Sol price reduction we reported on Saturday. For your company: if you have a list of small integrations that haven't been done for two years because "it doesn't pay to hire someone"—the connector between CRM and billing, the report someone manually copies every Monday—that list just changed price. The sensible way to take advantage of this is not to open an open bar: take one of those tasks, write the specification of what needs to happen and what should never happen, and let the agent work against it with human review at the delivery point. And a warning that is already being seen: the savings are real, but so is the dependency, so demand that what is generated is documented and executable by your team without the agent present. Source

Your agent can already pay, but no one can prove you authorized it

Fortune yesterday pointed to the gap that opens when an agent makes a purchase or a transaction that spans several companies: each system keeps its own piece—the merchant records the order, the bank the charge, your platform the session—and none link that charge to the task you assigned to the agent. Google's Agent Payment Protocol records the limits approved by the user and the information shown to each participant, but it doesn't decide who is liable or how long anything should be kept. In parallel, Senator Mark Warner registered the AI AGENT Act on July 21, which would require agents with mandates to keep real-time records of what they do for the user and tasks NIST with technical standards; its initial scope is an organization's internal agents and leaves for later the worst case, the agent crossing company borders. For your company: while this is being sorted out, it's up to you to set up the chain of custody, and it's cheap if you do it from day one. Three things: every task you launch to an agent must have its own identifier that travels in all the calls it generates, including those to external providers; that the authorization is written with its limit—how much, for what, until when—and not implicit in a permanent permission; and that the log is tamper-proof, meaning not even the agent itself can rewrite it. If tomorrow you dispute a charge of forty euros or forty thousand, the only thing that will serve you is being able to show who requested it, with what limit, and what the agent was shown before accepting. Source

What to watch tomorrow?

Nvidia presents results tonight, around 10:20 PM CET: the consensus expects about $92 billion in quarterly revenue and a forecast near $104 billion for the next, and what will really be watched is that forecast and the launch of the Vera Rubin platform. It is the thermometer of whether the money sustaining the entire chain—and with it the price of the tokens you pay—continues to flow at the same pace.

Watch the 1-minute video