self.md radar — 2026-09-24
the task continues after the window closes
A suspended agent run, a WAL checkpoint backlog, a warm introduction. The material gets specific quickly: state needs an owner, recording needs a recovery path, and the first plausible social choice may need to be dropped.
1. Google’s Agent Executor names the machinery behind a resumed run
Agent Executor (AX) is Google’s new Apache-2.0 distributed runtime for agents and harnesses. Its README puts isolated environments, suspendable and resumable images, a single controller, an event log and snapshots on the same diagram. That is a more useful admission than another claim about autonomy: an interrupted task has state somewhere, and somebody has to decide which writer gets to continue it.
The repository is explicit about the bill. AX is in active early development, expects major breaking changes before a stable release, and has paused external pull requests while it changes the core. A durable log is not a guarantee that the next action is right. It does at least leave a place to inspect why the runtime thinks it can continue.
reading: Google Agent Executor repository
2. Screenpipe’s new release spends its attention on recovery ownership
Screenpipe App v2.7.66 fixes several ways local recording can fall apart: a WAL checkpoint backlog, a Mac hostname change, shared writer pools, a storage-migration parity-scan livelock, and restart requests dropped during onboarding. The release also calls out preserving migration evidence through recording recovery.
None of that is a feature-tour line. It is the part that decides whether a machine remembers the work that already happened after an update, a restart or a renamed computer. The release cannot prove every data path survives in practice, but it makes the failure list legible enough to test.
reading: Screenpipe App v2.7.66 release
3. ReAdapt gives an agent one chance to abandon the obvious social move
The ReAdapt paper starts from a familiar mistake: an agent sees a salient post or a seemingly suitable contact and treats that as enough. Its synthetic benchmark has 500 social worlds and 1,000 queries; in about 53% of them, the surface-obvious option conflicts with the relationship-grounded answer.
ReAdapt adds an explicit state for goals, beliefs, relationships, norms and disclosure. After a tool observation, it must choose whether to continue, switch, abandon or clarify before taking the next action. On the paper’s 150-query-per-task evaluation subset, warm-introduction accuracy rises from 37% to 51%, while reaction selection rises from 69% to 77%, using the same model and tool interface.
Those are results inside a constructed world, not a licence for software to manage real relationships. Still, the mechanism is clean: retrieving another fact about people is useless if the system has no place to let that fact change its mind.
reading: ReAdapt paper on arXiv