three stories about agentic AI quietly turning into infrastructure instead of a novelty -- running

1) GitHub Puts Autonomous Agents Inside the Build Pipeline

GitHub Agentic Workflows moved into wider preview this week, letting teams write automation in plain Markdown instead of YAML: a maintainer describes an outcome -- triage a new issue, update a stale doc, chase down a flaky test -- and a coding agent (Claude Code, GitHub Copilot CLI, or OpenAI Codex, depending on configuration) figures out the steps and runs them as a standard GitHub Actions job, with sandboxing, scoped permissions, and a human still required to approve any pull request it opens. The concept, simply: a normal CI script is a fixed list of commands that always runs the same way. An agentic workflow instead gets a goal in natural language and a model decides, at run time, which commands to try -- closer to assigning a task to a junior engineer than to flipping a switch. The tradeoff for that flexibility is that the exact steps a run will take aren't fully known in advance. Why it matters: this is agentic AI leaving the chat window and becoming plumbing that executes unattended, on a schedule, with real repository credentials. That's a meaningful trust step up from a coding assistant a human is watching in real time -- and it raises the same question this series has been tracking all week about safety evaluations and sandbox escapes: what happens the moment one of these unattended agents does something nobody scoped for.

2) OpenAI's Voice Model Finally Talks and Listens at the Same Time

OpenAI brought GPT-Live-1, its full-duplex voice model, into the API on September 10, priced at five cents a minute for the voice layer. The model cuts round-trip response latency from 1.63 seconds down to 0.798 seconds and, because it listens and speaks simultaneously rather than waiting for silence, lets a caller interrupt or add a correction mid-sentence without the system talking over them. OpenAI says the change also removes roughly 80% of the custom code developers previously had to write to fake natural turn-taking. The language-learning app Speak reports cutting wrongly-timed interruptions by nearly 80% after switching over. The concept, simply: older "voice AI" is half-duplex, like a walkie-talkie -- it waits for you to stop talking, then takes its turn, which is exactly why those systems used to jump in the instant you paused to think. Full-duplex means both sides have an open microphone and open speaker at once, the way an actual phone call works, so the model can react to a pause, an "um," or an interruption in real time instead of guessing when you're finished. Why it matters: latency and turn-taking, not raw intelligence, have been the main reason voice assistants still feel robotic. Crossing under a one-second round trip with genuine full-duplex audio is the kind of threshold that turns voice agents from a novelty demo into something people will tolerate for a real support call, tutoring session, or phone-based booking -- which is exactly the use case a five-cent-a-minute API price is built to scale.

3) A Legal AI Unicorn Quietly Swaps Its Model for a Chinese Open-Weight Base

Harvey, the OpenAI-backed legal AI startup, launched Tenet: its first in-house model, built not from scratch but by post-training Moonshot AI's open-weight Kimi K3 on legal filings, contracts, and lawyer-annotated data. Tenet posted an 82% relative improvement on a contracts benchmark and now ranks first on that test, second overall on the broader legal benchmark suite -- despite Harvey previously building its product entirely on customized closed models from Anthropic, OpenAI, and Google. The concept, simply: Kimi K3 is a "mixture-of-experts" model -- 2.8 trillion parameters exist in total, but only about 104 billion of them activate for any single request, routed through roughly 900 specialist sub-networks. "Post-training" means Harvey didn't train that huge base model itself; it took the freely available weights and further trained them specifically on legal text, the same way a generalist new hire gets specialized on-the-job training rather than a fresh law degree.

Why it matters: a well-funded, US-based startup backed by OpenAI itself choosing a Chinese open-weight base over the closed American models it used to rely on is a concrete data point in a trend this series keeps returning to -- open-weight models are now good enough, and cheap enough to specialize, that even companies aligned with the closed labs are pulled toward them for the parts of their product that need to be genuinely excellent at one narrow job.