self.md radar — 2026-10-20
what gets to send the alert
One source counts missed handoffs in an apartment-leasing agent. Another sets a six-day clock for a missing web developer. Microsoft’s Scope makes the first kind of test repeatable in a stack somebody has to run.
1. The score needs a cutoff before it can page anyone
Hamel Husain and Antaripa Saha tested decision models on 100 traces from an apartment-leasing agent. The traces included failures such as a booking that did not happen and a handoff to a person that never came. They compared each model’s failure probability with human-reviewed pass/fail labels, then added API cost from token use.
The article does not treat an overall F1 score as the alert setting. For an automated alert, it uses 80% precision and 80% recall as a minimum example, then inspects the thresholds where both hold. Before a score can trigger a human, somebody has to decide whether a false alarm or a missed failure is the worse bill.
reading: Hamel Husain and Antaripa Saha, “How To Use Decision Models for Evals”
2. Scope puts coding-agent evals in a stack you run
Microsoft has open-sourced Scope, a platform for repeatable coding-agent experiments. It has a portal, CLI, REST API, agent workers, an evaluation service, and a Docker Compose quick start. The same scenarios can be submitted from a script or pipeline, then inspected as an experiment instead of recalled as a demo.
Scope is not a hosted service bundled with the release. Its operator chooses the connected services and manages experiment data, credentials, agents, and model-provider access. That list belongs in the evaluation plan, not under the benchmark chart.
reading: Microsoft Open Source, “Introducing Scope: Open-source evaluation for AI coding agents”
3. A six-day timer for the person who runs the domains
Death Plan Switch is a small self-hosted dead man’s switch for independent web developers. It sends a weekly check-in email; after 24 and 48 hours it reminds the operator, at 72 hours it asks next of kin, and at 144 hours it sends prepared messages to other contacts. It runs on PHP 8.1 or newer and does not require a database.
Opening the check-in link is ignored because mail scanners open links too. The operator has to press the button. That rule stops an automated system from falsely certifying that a human is still around to hold the keys.
reading: Death Plan Switch
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Browser agents — where browser automation helps, where it lies, and how to keep it observable