An agent that installs the app and looks at the screen
The capture loop that makes an on-device claim checkable: force-stop, install, launch, capture, hash. Why the order is load-bearing and what it cannot see.
Agents
Driving AI agents to software that is verified on a real device, not just compiled.
27 posts
Put the control flow in code, not in the prompt
Sequencing belongs in a script and judgement in the model. One routine commit cost 180k tokens as reasoning and seconds as bash. Here is the line.
Where to put the human
Gate placement for autonomous agents. Reversibility is the criterion, not confidence, and the consent surface matters as much as the question.
MCP, or a shell script? The honest comparison
Decided on concrete cases rather than preference. Where a protocol server earns its keep, where a script wins, and the failure that shows up unattended.
One level of fan-out, and why that is enough
A subagent cannot delegate, so the top session is the only scheduler. Designing around that, and why a barrier between stages costs more than it looks.
A gate that fixes what it finds
An auto-healing loop must fix the source, never the symptom, and must never edit the thing that found the problem. Confidence floors and bounded rounds.
Skills, tools, and the difference that matters
Progressive disclosure across four tiers: what must be loaded to decide, what loads to execute, and what should never load unless someone asks for it.
Stop parsing your AI's prose. Make it return a schema you can assert on.
Schema-forced returns move validation to the tool call, so the model retries instead of the caller parsing. What that buys, and the field that lied in the right format.
What to log when the actor is a model
Agent telemetry that survives the session: log the measurement rather than the verdict, and build a schema with no slot for the things that must not travel.
One tool call can destroy your agent's context window
What a tool returns is a budget decision. Measured savings from scoping one payload, and why a thin result is more dangerous than a fat one.
Teaching an agent when to stop
Enumerated halt classes make stopping checkable; a mood does not. The forbidden reasons that stall a drive, and the undeclared halt I made anyway.
Asking AI to check its own work does not work. Ask it to disprove it.
A verifier asked to confirm will confirm. Inverting the instruction to 'try to prove this wrong' found bugs that three rounds of review had passed, including one in the verifier itself.