Your agent fixes and proves it

Hey there 👋

I read the GitHub changelog twice this morning before the important line registered. The security agent is generally available now. The detail I want you to catch sits a sentence below that: before it opens a pull request, the agent reruns CodeQL against the fix it wrote to confirm the alert is closed. It checks its work before a person ever looks.

That is the thread through today. The frontier went open and local in a single day, Nvidia and Meta both put frontier-class models on Hugging Face, and xAI shipped a voice model you can deploy this afternoon. Under all of it sits one small idea: an agent you can trust is one that shows what it did.

- Ron

The Big Thing

GitHub's security agent writes the fix, then proves it closed the hole

GitHub pushed two code-security features to general availability yesterday, and together they change what the security backlog feels like. Agentic Autofix now takes a code-scanning alert, works through your codebase the way a developer would, and writes a fix. Then it does the step most AI code tools skip: it reruns CodeQL against the fix it wrote to confirm the alert is closed, and only then opens a draft pull request with a summary of the change and the validation steps it took. You get a reviewable diff that already passed the same scanner that raised the flag.

The companion release, AI Scan for pull requests, widens the net. CodeQL has never covered a long list of languages that real codebases are full of, so AI Scan adds findings for PHP, Shell, Terraform and HCL, Dockerfiles, JSP, and Blazor. These findings are advisory by design. They do not fail your checks or block a merge, they show up labeled so a reviewer can weigh them. And AI Scan needs no Copilot license to switch on.

Read the billing before you flip it. Agentic Autofix needs a Copilot billing method defined, and both features draw down Code Security AI Credits, with Autofix also spending Actions minutes on every run. Public repositories are covered for organizations and personal accounts; private and internal repos require GitHub Code Security or Advanced Security. So set a credit budget before you point the agent at a few hundred open alerts.

Ship it? DEPLOY, in that order. Turn AI Scan on first, since it is advisory and adds no risk, and let it surface problems in the languages your scanner has been blind to. Then pilot Agentic Autofix on a handful of low-severity alerts with a human reading every PR, and watch whether the CodeQL re-check holds up on your code. What makes this matter past the feature list is the pattern it ships by default: the agent hands you a diff that already passed the scanner, with its validation steps shown, so your review starts from evidence. Copy that habit into every agent you run.

Sources: GitHub: AI Scan for PRs GA; GitHub: Agentic autofix GA.

Tour de Headlines

🧠 Nvidia open-weighted a multimodal model built to be post-trained. Nemotron 3.5 Super VL is a 120B multimodal model under the OpenMDW license, which lets you use it commercially, fine-tune it, and redistribute it, with the weights on Hugging Face today. Two smaller Bolt variants come with it: a 74B-A9B that squeezes into 24GB of memory, and a 4GB E4B due October 23 for phones and edge devices. Keep your skepticism pointed at the benchmark chart. Nvidia shows post-trained versions beating Claude Opus 5 and GPT-6.1 Sol on five vertical tasks, but those are customized models measured on specific jobs after a customer fine-tuned them, the test protocol is unpublished, and the base model's general score sits below several open peers. Count those wins as Nvidia's until someone outside the company reproduces them. Verdict: PILOT if you have a narrow task and data that cannot leave your walls.

🎙️ xAI shipped a voice model you can deploy this afternoon. Grok Voice 3.0 is a speech-to-speech model now at the top of the Artificial Analysis speech index, it answers in under a second, and the price held at eight cents a minute. If you already call grok-voice-latest, you were upgraded automatically. One honest caveat keeps it from being a clean sweep: OpenAI's GPT-Live-1 still scores higher on the noisy support-call test, so a tidy voicebot and a messy contact center will not rank these two the same way. Verdict: DEPLOY if you build voice agents, but run it against your own recordings before you cut over, because the benchmark that counts is the one made of your calls.

💻 Meta open-weighted an 816B coder you can run on two Mac Studios. Following up on Muse Wonder, the closed frontier model we covered October 13, Meta released its open sibling: Muse Spark Open, 816B total and 42B active, with a 1M-token context window and a license that allows limited commercial use. It is built for long-horizon coding and agentic work. The flex is where it runs. It went live the same day across seven hosted providers, from Groq to Together, and it runs locally on a pair of M5 Ultra Mac Studios through Apple's MLX framework. Meta is betting on reach, putting a frontier coder on every cloud and on your desk at once, the same morning Nvidia made the same move. Verdict: TRY it through a hosted endpoint, and PILOT the local path if your code cannot leave the building.

Tool of the Day

🛠️ AWS AI-DLC

Make your coding agent follow a real delivery process instead of improvising into your repo.

If you let an agent write into a real codebase, the cheapest safety upgrade you can make this week is a process it has to follow. AI-DLC, from AWS Labs, is a set of open-source steering rules that wrap a coding agent in a structured delivery lifecycle: five phases, thirty-three stages, approval gates at the seams, and an audit trail of what happened. It runs inside the agents people already use, Claude Code, Kiro, Codex, Cursor, opencode, and Copilot, and it ships under the permissive MIT-0 license. It fits today's theme cleanly. A coding agent that has to check in at each stage is a coding agent whose work you can review and trust. Install it on one repo and run a single feature through the workflow before you decide.

AI-DLC on GitHub →

Worth a Click

• The token bill comes due. Accenture now funds just $29 a month of Claude per employee, with overages charged to a project. Uber and Walmart capped the same week. Audit your own per-seat AI burn before finance does.
• Where to run agent code. Modal raised $750M at a $15.75B valuation, and the deployable part is VM Sandboxes going GA, roughly 3 billion spun up and scaling to a million concurrent. A production answer to isolating the code your agent writes.
• A private agent on your laptop. Liquid AI's LFM2.5-24B-A2B is an open-weight 24B/2B-active model that runs at 100-plus tokens a second on a laptop and beats a larger Qwen on agentic automation. Use it as a fast local orchestrator, since bigger models still win on heavy knowledge and coding.

One habit runs under the whole issue, and I will put a stake on it. The most useful thing GitHub shipped was not a smarter model. It was an agent that reran the test on the code it wrote before asking a person to look. That is the bar to hold every agent to this week: a reran check, a validated diff, a closed alert, a passing build. The agent should show its work, the way you would expect a junior engineer to show a passing test before you merge their branch.

The open and local model drops are the other half of the day. If your data cannot go to a hosted API, Nvidia, Meta, and Liquid just handed you frontier-class options you can run on your own hardware, which means the proof-of-work discipline now reaches the teams who could never send their code out in the first place. So the move this week is simple: treat your agents like a junior engineer you like but do not yet fully trust, and ask for the evidence before you merge. The teams that win Q4 will be the ones that make every agent show the receipts.

- Ron