AI systems that take action, and the tools and safeguards around them.
Anthropic's free, opt-in scanner offers open-source projects recurring AI vulnerability reports. Its value depends on whether maintainers can turn those findings into verified fixes without overwhelming the rest of their work.
Claude's Google Workspace beta lets it edit documents, spreadsheets, and slides where teams already work. That changes how revisions should be scoped, reviewed, and handed back to the people responsible for the files.
Together AI introduced Link to connect existing coding tools to hosted open models. Its session-level routing promises a simpler switch, while meaningful cost comparisons still require equivalent tasks and accepted results.
Microsoft and Hugging Face presented ThinkingBox, evaluating business workflows through final database state and repeated trials. The approach makes a useful distinction between plausible tool activity and a correctly completed transaction.
Claude Code introduced TypeScript mods for behavior and interface customization. The flexibility is useful for teams with specialized workflows, but local extensions need versioning, clear purpose, and a simple way to disable them.
H Company released Holo4 models for computer use across multiple interfaces. The ability to switch between a screen and an API is promising, but consistent state and recovery become central engineering concerns.
GitHub Security Lab released an agent workflow for setting up and running C and C++ fuzzing. Its value depends on reaching meaningful program behavior and turning crashes into reproducible engineering work.
Google added local model support to the Antigravity SDK through LiteRT and Gemma. Offline execution is useful, but hybrid workflows still need an explicit account of what leaves the device.
