A pager and a phone sound identical on first glance. One is an inbound paramedic crew calling the emergency department with a STEMI ECG en route. The other is the ward, calling because a patient has increasing pain. Both say: act now. And yet they land differently in me — and I have been thinking about why.
Triage is time allocation, not sorting
The ABCDE framework in emergency medicine reads at first like a priority list. It is not. It is a scaffold that allocates attention independently of the loudest symptom: airway first, then breathing, then circulation, then neurology, then everything else. A screaming patient in visible pain instinctively pulls attention — but the quiet, pale one next to them may die faster.
In agent management in my homelab I have noticed a similar gravity: you respond to the loudest alert rather than the most important one. When an agent fails loudly — exception trace in Grafana, half a chat history drowning in retry loops — the eye jumps there. But a retrieval index that silently returns stale embeddings can quietly poison every subsequent agent response, while the exception is just an overflowing tool call that a fallback handler already absorbs.
ABCDE works because it prioritises independently of emotional weight. The same discipline in an agent incident would be a fixed order: model endpoint first (is it responding at all?), then context pipeline (is the right information arriving?), then tool chain (are calls executing cleanly?), then agent logic (prompt, routing, state). Only once each of these foundations is upright is the loud blinking trace allowed to be looked at.
Severity is not the same as drama
In the clinic I distinguish between vitally threatened, acutely in need of control and painful but stable. All three sound dramatic. But the vitally threatened patient needs seconds, the second minutes, the third hours.
The homelab has the same distinction: a degraded agent (higher latency, lower answer quality, more retries) is something different from a failed agent, which is again something different from a compromised system — for example an agent that successfully ran a prompt injection against itself and is starting to exfiltrate data. Only the last justifies getting up at 3 a.m. The first two can wait until morning, if the runbook entry exists.
My personal mistake in my first months with agent alerts was treating all three as equally urgent. That was not conscientious — that was burnout in preparation.
Checklists are not for those who can’t do it
For a STEMI I always go through the same checklist and the same standards. I know them by heart — by now they are internalised, and I have taught them many times. I still look at the list every time. Not because I don’t know it, but because alertness is not a constant. At three in the morning, at the end of a long shift, even routine slips.
Agent deploy checklists exist for exactly the same reason. A new prompt, a changed tool schema, one extra retrieval hop — each individual step is trivial, the interplay is not. The best time for a productive mistake is when everything feels familiar and you skip the eval suite “just this once”.
In both worlds the checklist helps precisely the seasoned people who are most likely to think they no longer need one.
Where the analogy ends
People are not agents. Clinical triage carries an emotional toll — relatives wait, patients are afraid, team dynamics wobble. That overhead is absent in the homelab incident, and that really is a relief. You can debug a system coldly without hurting it in the process.
Conversely, the clinic teaches something that engineering often forgets: not to personify technical problems. An agent that hallucinates is not a culprit. A failing deploy is not a character trait. The ABCDE posture in a system incident also rules out hunting for “who broke it” — and replaces that instinct with “what is missing in the setup so this can’t happen again”.
This is the kind of crossover thinking I will probably write about more often here. Not because medicine and agent management belong together topically, but because the mental muscles that both train turn out, more often than I expected, to be the same ones.