GitHub Security Lab released a fuzzing taskflow agent this week, and it’s actually generating working fuzz harnesses. Not suggestions. Not boilerplate. Full fuzzing implementations that find real bugs.

The agent is built on GitHub’s Taskflow Agent framework and handles the entire process: analyzing codebases, identifying fuzzing targets, writing harness code, and running the tests. It works with libFuzzer and integrates with OSS-Fuzz, the project that’s found over 10,000 bugs in open source software.

If you maintain a C or C++ project, this matters. Fuzzing is effective but tedious to set up. Most projects don’t bother. An agent that writes the harness code removes the main friction point. GitHub is positioning this as a way to get more projects into OSS-Fuzz’s continuous fuzzing pipeline.

The code is on GitHub. The agent is part of Security Lab’s broader AI framework, which already handles other security tasks. This is the first fuzzing-specific implementation.

Datasette adds observability, inches toward 1.0

Simon Willison released Datasette 1.0a41 with OpenTelemetry support, courtesy of Alec Garcia. You can now instrument your Datasette deployments with standard observability tooling.

This matters if you’re running Datasette in production. OpenTelemetry gives you tracing and metrics using the same libraries everything else uses. You can pipe telemetry to Honeycomb, Datadog, whatever you’re already running.

Willison also refactored Datasette’s modal dialogs into a single Web Component that’s now documented for plugin authors. If you’ve written a Datasette plugin, you can use the same modal pattern the core uses instead of rolling your own.

Datasette has been in alpha for years. These aren’t flashy features, but they’re the kind of production-readiness work that precedes a 1.0 release. Observability support especially signals that Datasette is getting serious about operational deployments beyond prototypes and side projects.

The release also includes the usual bug fixes and refinements. Check the changelog for details.

Ando wants to be Slack for humans and agents

Ando launched this week with a team messaging app where AI agents get their own identities and inboxes. Not bots that respond to @mentions. Full participants with presence, message history, and the ability to jump into conversations.

The pitch: if agents are doing real work, they need real infrastructure. Ando gives each agent an inbox, lets them initiate conversations, and tracks their work the same way you’d track a teammate’s. You can @mention them, they can @mention you, they can @mention each other.

This is either prescient or premature depending on how fast agents actually become coworkers. Slack added AI features. Microsoft added Copilot to Teams. Ando is betting those are bolted-on features and there’s room for a messaging platform built agent-first from the start.

No pricing announced yet. The company didn’t share much about how agents authenticate, how they’re provisioned, or what happens when an agent screws up and spams a channel. Those details matter more than the concept.

If you’re building agent workflows and finding Slack limiting, Ando might be worth watching. If you’re still figuring out whether agents are useful at all, this probably isn’t your problem yet.

Everything else

commit-rewriter 0.2 shipped with support for branches other than default. Run uvx commit-rewriter --branch other to rewrite commits on a different branch. Small tool, useful if you’re automating Git history cleanup.

Google is testing Gemini phone calls for Pixel 11 owners with Gemini subscriptions. The AI will call businesses for you. US only, subscription required, Pixel 11 only. This is extremely limited access for something that’s mostly a demo of what voice agents might do eventually.

Google Photos launched a virtual closet feature on Android and iOS. It’s the Clueless-inspired wardrobe organizer that started rolling out to Android users in June. AI categorizes your clothing photos into a browsable closet. Cute idea, narrow use case, probably not your workflow unless you photograph your clothes regularly.

PrismML is putting tiny LLMs on Qualcomm-powered smart glasses. The goal is on-device inference that uses existing hardware better. Open-weight models, local processing, the whole edge AI pitch. Smart glasses are still looking for a reason to exist, but if they do find one, on-device AI will probably be part of it.