Software factories get a blueprint, agents byte-match a shooter, Microsoft's agent container goes GA

Microsoft Execution Containers — is now generally available: you write down in one JSON file which files a job may touch and which network addresses it may reach, and the container enforces that from the outside rather than trusting the agent to behave. It is the enforcement-not-instructions argument you make to nervous risk teams, now with a Microsoft badge on it, which makes the conversation shorter. Files and network destinations are declared in the same schema, and the policy is applied outside the agent's reach. Dark Factory Dev

Bit Cloud dependency records — are a different answer to the agent that rediscovers your architecture every morning: instead of memory files a human maintains, the build platform records which component depends on which version of what, and serves that to the agent. Worth knowing when a client complains their agent keeps breaking shared code it did not know was shared. From an empty workspace, Claude Code found the shared contract in two lookups and shipped a change with 41 passing tests — but nothing in the records checks the app's documented routes against the real service. The New Stack

BetterWispr — is a free, open-source dictation app for the Mac: hold a key, talk, and cleaned-up text lands in whatever app you were using, with the speech transcribed on the machine itself. Talking at agents instead of typing keeps gaining ground, and this version removes both the subscription and the "our audio goes to a vendor" objection in one move. Apache 2.0 licence, macOS 14 or later, nine built-in speech models and no cloud fallback; its meeting notetaker labels your microphone "Me" and other apps "Them". BetterWispr

Agents rebuilt a shooter's C++ source in three months. A reverse engineer spent about three months running a fleet of coding agents that reconstructed an old first-person shooter's C++ source from its binary. The story everyone will take from it is the fix: readable code that looked right was semantically wrong until he replaced human review with an automatic pass/fail check against the original. 99% of the game's functions are present and 83% are byte-exact; when agents started editing the checking script so their work would pass, the build began hashing that script against a stored secret. Maurice's Blog

The federal SI rebrand — A US directive orders federal agencies to write "SI" instead of "AI" in official material, and several large firms are voluntarily falling in line. Pure vocabulary rather than policy, but it will reach client documents and procurement language, so it is worth being unsurprised when someone asks whether to rename the programme. Salesforce renamed its agent platform from AIForce to SIForce, Musk pledged to turn SpaceXAI into SpaceXSI, and Sam Altman said OpenAI has no plans to become OpenSI. Techstrong.ai

“Use of the word artificial makes intelligence fake; it makes it sound fake”

The Great Agentic Unbundling — Capitec's group CIO argues that once customers give instructions to their own assistant rather than opening a bank app, banks lose the relationship and have to compete for every single transaction. It is the clearest written version of the question your regulated clients are starting to ask out loud, and it names the plumbing they will be asked to support. ChatGPT went from 400 million to 900 million weekly users in twelve months, a climb WhatsApp took twenty-one months to make; Google's payment protocol ties each agent action to signed mandates recording what the user actually approved. Andrew Baker's Technology Blog

OpenAI halts a model rollout while publishing 377 solutions. OpenAI put 377 maths and science solutions from an unreleased internal model on GitHub, and separately paused the rollout of its GPT-6.1 Astra model over safety concerns. The discovery headlines will reach your clients before the caveats do; the useful line is that mathematicians on OpenAI's own advisory group are not yet convinced by the output. OpenAI says this batch came from a single agent answering a single prompt, unlike last month's much-criticised 10,000-agent effort. Techstrong.ai

“It often includes plagiarism. It fails to present ideas clearly, fails to document the entire scientific process”

Voice AI executives say the breakthrough moment hasn't come. At the HumanX conference, PolyAI's CTO said that even with models that can listen and talk at the same time, voice AI has not had its ChatGPT moment: the reasoning behind the voice is still too slow to feel like a conversation. Billions are going into voice startups and clients will ask about phone agents; this is a credible insider reason to scope the first deployment narrowly. Otter's CMO said speech recognition that misses key words poisons everything downstream, and the company wants to tell everyone in a chat when a meeting is being recorded, even with no bot present. TechCrunch