Triage
A transport-agnostic AI agent that autonomously diagnoses faulty computers.
Triage is an AI agent that diagnoses faulty computers on its own — running system checks, applying safe software fixes, and escalating hardware or high-risk issues to a human operator. It is built around a modular action pipeline with explicit safety boundaries: the agent may only execute repairs that fall inside a defined low-risk scope, and anything beyond that scope becomes a report for a person rather than a command.
triage run --scope low-risk
✓ disk SMART ok · 41% used
✓ memory no ECC errors
! services 2 failed units → auto-repair
⚠ battery wear 38% → escalate: hardware
report ▸ 3 fixed · 1 escalated
// engineering notes
Decisions behind it
01
Safety boundaries are structural
Low-risk repairs are the only actions the agent can reach. Hardware faults and destructive operations are not gated behind a prompt instruction — they are outside the action set entirely, and route to a human operator instead.
02
Modular action pipeline
Each diagnostic and repair is an independent, testable action with declared inputs, effects, and risk level. Adding a new check means adding a module, not editing a monolith.
03
Transport-agnostic
The agent core is decoupled from how it is reached, so the same diagnostic engine can be driven locally or over a remote channel without changing its logic.
04
Tool-use over free-form generation
The model chooses tools and reads their real output rather than narrating what it thinks a machine would say. Every conclusion traces back to a command that actually ran.