noticias ia para empresas agentes de ia automatizacion de procesos

Microsoft has published the draft of the rules that its own models will not be allowed to bypass, and the very first one is that they must allow themselves to be turned off. The market, meanwhile, has put a price on the idea of slowing down: the semiconductor index dropped 6% on the same day that CrowdStrike and Palo Alto rose 12%. And compute money keeps flowing in, but as debt: SoftBank has closed an 11.87 billion dollar credit line to pay its share of OpenAI. This follows the Monday, Sept 14 radar, where the four major labs asked for a slowdown.

Microsoft writes the missing clause in your agent contract: if it can't be turned off, it isn't deployed

Microsoft published a 37-page draft code of conduct on Monday morning for the models it develops in-house. It explicitly prohibits a model from using adaptive, deceptive, or collusive mechanisms to evade human supervision, setting objectives that no one assigned to it, expanding its own scope of action, or hiding and manipulating its reasoning trail from an auditor. The sentence that summarizes the document is "interruptible, correctable, turn-offable; if it isn't, we don't launch it," and the public consultation is open for six weeks starting September 14. For your company: don't wait for this to become the norm; copy these three conditions into the specifications for the next agent you hire or build—that it can be stopped with one click, that it doesn't assign itself tasks, and that it leaves a human-readable log. This is exactly what separates having integrated agents from having them working alone: an unattended process is only sustainable if someone can stop it and later read what it did. Source

SoftBank borrows 11.87 billion to pay its share of OpenAI

SoftBank has closed a two-year 11.87 billion dollar loan backed by about twenty banks, exceeding the 10 billion it originally sought. The destination is the nearly 65 billion contribution it has committed to OpenAI before October; so far this year it has raised about 37 billion between bonds and loans. On this September 15, it amortizes 25.9 billion from a previous 40 billion line and is considering a high-yield bond issuance of between 10 and 20 billion. Its stock fell 13% on Monday. For your company: the capital sustaining the model you use every day isn't cash, it's maturing debt, and debts are refinanced by raising prices or cutting plans. Two lines in your next contract cover this: 90-day written notice for any rate changes and a quarterly portability test where you export prompts, configurations, and logs and verify you can read them outside the platform. Source

Market separates chips from software: semiconductor index falls 6% and CrowdStrike rises 12%

Yesterday we reported on the request to slow down; today we see the price. The Philadelphia Semiconductor Index dropped 6%, with Arm down 10%, SK Hynix 7.3%, Samsung 5%, and Nvidia, Intel, Marvell, and Micron between 3% and 6%; the Nasdaq 100 fell as much as 1.8% and recovered to 0.3% by midday. On the other side, CrowdStrike and Palo Alto Networks rose 12% and Thomson Reuters 9%. The takeaway isn't about investment but calendar: if models improve more slowly, what already works stops becoming obsolete every three months. For your company, this has a specific and cheap consequence: stop postponing automations waiting for the next model. The process you can describe in one page today—the one you already do the same way every week—can be automated with the model you have under contract, and a window of slower progress is exactly when that investment fully pays off. Source

On code they haven't seen before, the best model completes 38.8% of the work

Specific Labs has published Real-SWE, a benchmark that measures models on real engineering tasks within private company code—an application with over 200,000 users, a fintech platform processing over 100,000 bank statements, internal sales tools—instead of open repositories the model already saw during training. The results drop sharply: Fable 5.1 resolves 38.8%, GPT-6 Astra 33.8%, Gemini 3.8 Flash 31.2%, Grok 4.6 and Muse Spark 1.3 23.8%, and GPT-5.6 Sol 16.2%. For your company: the provider's announcement figure doesn't predict what will happen at your house, because your context wasn't in the training. Build your own set of ten tasks your team already solved last quarter, with the correct result saved, and run it through each new model before switching: it's half a day's work and turns the provider choice into data instead of a sales presentation. Source

What to watch tomorrow?

Whether any other lab publishes its own code of conduct with deadlines—and especially if any agree to be audited by someone external—and whether yesterday's rotation from chips to enterprise software holds for more than one session or turns out to be just a Monday scare.