self.md radar — 2026-09-23
an agent gets a family account; another stops spending turns on a waiting room
Google is giving a household assistant its own account. A new runtime is making tool waits somebody else’s problem. A coding-agent guard is looking backward through session traffic. The interesting parts are literal: who can share, when the model is allowed to wait, and which values a detector refuses to pretend it can see.
1. CC becomes a household account
Google’s CC is no longer framed as one person’s morning brief. It is now a household agent: up to six people can collaborate with it through a verified Google Account that belongs to CC itself.
Each person chooses what to share. Google says CC only replies to group members and cannot act or share outside the group without permission. Behind that boundary sits an isolated cloud computer running the Antigravity harness; Google’s examples include pre-filling a class-registration PDF and checking drive times between back-to-back activities.
The previous CC was a personal assistant. This version has to hold a shared permissions boundary before it can hold a family calendar.
reading: Google Labs: CC is an AI agent for families and groups
2. Unreal Agent moves tool waits off the model turn
Unreal Agent runs tool calls asynchronously. The model need not spend turns waiting, polling, or sending heartbeats; it can continue other work and receive results when they arrive.
Unreal Labs is publishing a Go library, a runner comparable to claude -p or codex exec, and a Harbor-compatible benchmark runner. It claims up to 40% lower cost than Codex and up to 20% lower than Pi across its production workloads and coding/science benchmarks. That is a vendor comparison, not a general law of agent economics.
The release makes scheduling part of the harness: a long-running tool does not have to occupy the next model turn just to say it is still running.
reading: Unreal Labs: Unreal Agent
3. Stroq checks what a past session sent
Stroq’s v0.14.0 release
adds stroq sent, which reads existing coding-agent session recordings and reports credentials that appeared in traffic to a model provider. The release deliberately calls this “sent,” not “leaked”: the record shows that a value was present in the route, not that anyone abused it.
The same release adds MCP masking for exact credentials and validated email, phone, IBAN, card and SSN values. Names and street addresses are still visible by design. It also adds stroq run --, which hardens Git before an agent’s startup git status; its stated test case is a repository whose core.fsmonitor would otherwise execute on that first command.
The release says stroq sent was tested over 230 sessions, covering 3,180 tool calls and 122 MB after two false-positive fixes. The scope is narrow enough to inspect.
reading: Stroq v0.14.0 release