Do we still need a human on-call rotation if an AI agent resolves most incidents autonomously?

The shift

Detecting, diagnosing, and remediating the common, pattern-matchable slice of incidents — restart the flaky service, roll back the bad deploy, scale the saturated pool, apply the known fix from the last ten times this alert fired — goes from scarce human interrupt-time to abundant, always-awake agent capacity that acts in seconds instead of paging someone at 3am.

The axioms

• Rotation exists to guarantee 24/7 coverage because incidents happen on no schedule and someone must always be reachable. Rests on availability being a human scheduling problem.
• An on-call engineer is needed because diagnosing and fixing an incident requires expertise embodied in a trained person. Rests on diagnostic and remediation skill being scarce and person-bound.
• Pager burden is spread across a rotation because no one person can sustain constant interrupt-readiness without burning out. Rests on human attention and rest being finite.
• A human has to be woken up because someone must authorize or execute the action that ends the incident. Rests on action-authority and accountability being human properties.
• Getting paged is how engineers build system intuition — the toil teaches the system. Rests on tacit operational knowledge being acquired only through lived exposure to failure.
• The org needs a human escalation path because some incidents have no precedent to match against. Rests on judgment under novel ambiguity being irreducibly human.
• Rotation depth (how many people, how many weeks between shifts) is sized to expected incident volume and MTTR targets. Rests on response capacity being bounded by the number of humans available to take a page.

Invalid axioms

1. An on-call engineer is needed because diagnosing and fixing an incident requires expertise embodied in a trained person. Correlating alerts, logs, traces, and deploy history against thousands of past incidents and applying the known fix is exactly the synthesis-plus-action work agentic AI now does continuously, without waiting for someone to wake up and load context. The habit-trap: rotations are still staffed and scheduled for the volume of routine, pattern-matched incidents, when that volume is precisely what's shrinking fastest.
2. Rotation depth is sized to expected incident volume and MTTR targets. If an agent closes most tickets before a human would have even acknowledged the page, the old sizing math — enough people to keep MTTR acceptable at a given incident rate — is solving for a rate that no longer reflects what actually needs a human. The habit-trap: teams keep the same rotation length and shift count calibrated to pre-agent incident volume instead of resizing around the residual, harder incidents that remain.
3. Getting paged is how engineers build system intuition. If an agent absorbs the bulk of routine pages, fewer engineers accumulate the 3am pattern-recognition that used to come free with volume. This isn't fully invalid yet — see NEW #2 — but the belief that rotation-as-training is the mechanism that builds judgment is already false for the pages an agent now intercepts before a human sees them.

Unchanged axioms

1. A human has to be woken up because someone must authorize or execute the action that ends the incident, and answer for it if it's wrong. An agent can restart a service or roll back a deploy, but when the action itself is ambiguous, irreversible, or touches customer data, money, or a physical system, someone accountable has to be reachable to say yes — or to own the outcome when the agent acted autonomously and it went wrong. Autonomy doesn't remove the liable party; it just changes when they get pulled in.
2. The org needs a human escalation path because some incidents have no precedent to match against. Agent remediation is confidence built on prior incidents, runbooks, and known failure modes. The genuinely novel failure — a new cascading interaction, a vendor outage with no playbook, a security incident that looks like an outage — is exactly where a confidently-wrong autonomous action does the most damage, and exactly where a human still has to reason from scratch under pressure.
3. Someone still has to verify that the agent's read of "resolved" is actually true. "Most incidents resolved autonomously" is a claim about a success rate, not a guarantee — an agent can mark something resolved because the symptom it was watching went away while the underlying cause didn't. A rotation that never gets paged has no chance to catch that; someone has to be positioned to notice when "autonomous resolution" quietly failed.
4. Physical and cross-boundary actions still require a human or a tightly scoped, human-authorized automation. Cutting over hardware in a colo, coordinating with a cloud provider's own incident response, calling a vendor, communicating with customers or regulators during an outage — these sit outside token generation and stay human-gated regardless of how good diagnosis gets.

New axioms

1. When an agent resolves an incident with no human in the loop, who finds out it happened and checks the resolution was sound? If the rotation only gets paged for what the agent can't handle, the team loses visibility into the majority of incidents by default — the escalation path can quietly become the only signal anyone has about system health, and it's a biased sample (only the hard cases).
2. If routine pages disappear, what replaces on-call as the mechanism that builds the judgment needed for the incidents that still reach a human? The residual rotation exists specifically for novel, high-stakes, unclear-precedent incidents — the hardest category — but the training ground that used to produce people capable of handling that category was the routine pages now gone.
3. Who is on the hook when the agent's autonomous action causes or worsens the outage, and how fast can a human intervene once it's already acting at machine speed? A human paged after the fact to review a decision that already executed is not the same safety net as a human who approved the action before it ran. The faster the agent acts, the shorter the window to catch it before consequences land.
4. Does a thinner, less-frequently-paged rotation atrophy the readiness of the humans still on it? A rotation that pages once a month instead of three times a week keeps people nominally on-call but with less recent hands-on context, less current mental model of what's actually running in prod, and a longer time-to-effective-response when they are needed.

Where it breaks

"Rotation depth is sized to incident volume" (invalid) collides with "who builds the judgment needed for the incidents that still reach a human" (new): teams shrink rotations because the agent absorbed most volume, but the incidents that survive to reach a human are the hardest, novel, highest-stakes ones — and the smaller, less-frequently-paged rotation is staffed by people who got less practice, not more, right when the remaining job got harder.

A second collision: "an on-call engineer is needed because diagnostic expertise is scarce" (invalid) meets "someone has to verify the agent's resolution was actually sound" (still holds/new): if nobody's being paged because the agent marked things resolved, the org has no built-in trigger to catch a false "resolved" — the same automation that eliminated the need for constant paging also eliminated the tripwire that used to force a human to look.

ai-axioms.vercel.app/axioms/engineering/sre-on-call-rotation-ai-axioms