Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

drivers, replays, executable explanations

computer-use drivers, exact replay verification, and executable Prolog policies point at agent evidence that can be rerun, inspected, or edited.

Published Updated Read 4 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-07-20

the interesting object today is not the agent’s answer. it is the thing left behind that another system can drive, replay, or edit.

one repo is turning computer-use into drivers and fleets. two papers separate “the reviewer was right” from “the system repaired itself.” another paper takes a neural policy and spits out runnable Prolog. nice little day for anyone tired of screenshots as evidence.

1. computer-use got a driver layer

sources:

what happened: trycua/cua pushed fresh code today around an MIT-licensed stack for “computer-use 2.0”: open-source drivers, cross-OS fleets, benchmarks, evaluation, and data generation. the README is not selling one magic agent. it splits the surface into Cua Drivers, Cua Sandbox, Cua Bench, and Lume macOS virtualization, with Windows sandbox and macOS virtualization sitting in the repo metadata too.

GitHub had the repo at 20,278 stars when the cache ran. that number is noisy, sure, but the shape is not: computer-use is turning into a substrate people can test against instead of a demo where an agent heroically clicks a browser for the camera.

why this matters: once the desktop becomes an action surface, the important question is who owns the driver, the sandbox, and the benchmark. self.md cares about that layer because personal agents do not get safer by being more charming in chat; they get safer when their hands are visible.

2. reviewers are being judged by uptake, not applause

sources:

what happened: one new ARC-AGI-3 paper compared four nested Codex-based agents: plain text, executable world model, scheduled simplification, and a fixed-interface verification treatment requiring exact replay of recorded observations. the verification variant ranked first in all four main model-effort settings, though it spent more resources, and a gpt-5.6-sol follow-up solved the public games with about 99% RHAE. the authors also leave the sober warning in place: the model postdates the public games, and held-out performance is still untested.

another paper tested 4,181 verifier-grounded Omni-MATH problems and found the awkward part of “reviewer agents.” the planner-executor-reviewer setup had a more precise reviewer than broadcast discussion, 0.861 versus 0.644, but its useful critiques were less likely to change the next candidate and produced weaker repair.

why this matters: agent evaluation is finally poking the joint between noticing an error and carrying that correction into the next move. a review step that cannot change the run is just a very educated spectator with a clipboard.

3. explainability got compiled into Prolog

sources:

what happened: a new arXiv paper asks whether a trained reinforcement-learning policy can be rewritten as an executable logic program. the pipeline freezes a PPO teacher, induces ordered rules, emits Prolog, then lets an expansion loop edit the rule base only when policy evaluation certifies a return increase.

that is a very different flavor of “explainable AI.” the paper claims four guarantees, reports a two-room key-and-door task with 16,944 reachable states where the expanded Prolog program reaches exact optimal return in every seed, and shows the ceiling on harder continuous-control tasks: roughly 97% CartPole return, weaker LunarLander, and an exponential cost when the boundary gets ugly.

why this matters: readable explanations are nice. executable explanations are meaner and more useful: a person can inspect them, a logic engine can run them, and an optimizer can argue with them without pretending the neural net has become a diary.

  • AnovaX local assistant — a small local-first desktop voice assistant with typed executors, whitelist/denylist safety, bounded nesting, and a phone mirror over local WiFi.
  • SaneProcess — AGENTS.md plus hooks, skills, MCP, scripts, stop conditions, and release checks for coding-agent work.
  • mustflow — a repository-local read/command/verification contract so agents do not borrow permission from vague context files.
  • NoWreck — a deterministic checker for whether an AI coding explanation matches the real diff.
  • r3 — local line-level review for agent-written diffs and docs, because chat is a terrible place to track five unresolved comments.

left on the table

computer-use agent-evaluation explainable-ai personal-ai