The strongest releases this week hardened the mechanics that keep an agent useful after the demo, not the mechanics that make the demo work.
FileSystem traversal events now reach RepoContext through @on_event, and a repo-defined Pydantic AI agent can run inside GitHub Agentic Workflows.
Evented plumbing plus a route to run the agent inside GitHub Actions, rather than another capability bolted onto the harness.
Windows sandbox runs can record blocked access and take a constrained escalation before any full bypass, and Vim mode is now generally available. This is a pre-release.
Sandbox policy is enforced inside the coding-agent loop rather than around it, so a blocked call becomes an observable event instead of a failure.
The patch rolls up desktop session controls, MCP authorization, cron scheduling and delivery fixes, and delegation reliability work.
A broad always-on runtime with session, authorization and scheduling in one place, rather than a narrow coding harness.
An open-source software factory that connects GitHub, Linear and Slack, moving issue intake, planning, implementation and PR review through configurable agent stages.
Mastra publishes its own alpha counts rather than only a headline range: 277 of 1,627 merged PRs authored (17.0%) and 222 of 778 issues closed (28.5%). The same post also claims 25-35% of PRs and 50-60% of issues; where a claim and its underlying table disagree, the table is the figure worth carrying.
Teams evaluating agents should now inspect event surfaces, sandbox policy, authorization, scheduling and recovery — not just model quality or tool count. The question has moved from what an agent can do to whether it stays governed while doing it.
