The 3 a.m. page is becoming an agent's shift. Here's what the engineers do instead.
On-call is quietly changing shape. Teams running agents against their telemetry report fewer human pages, longer uninterrupted sleep, and a new, harder job for the people who used to answer.
Ask a platform engineer what they'd change about their job and the answer hasn't moved in a decade: the pager. Being woken at 3 a.m. to restart a thing that will restart itself in four minutes anyway is the tax everyone pays for running production systems.
That tax is starting to fall. Not because systems got more reliable, but because something else is picking up first.
What's actually happening
In conversations with teams that have put agents on their operational data, a common pattern has emerged. The agent sits on the event stream the alerting system already reads: metrics, logs, deploy events, feature-flag changes. When an alert fires, the agent gets it first. It correlates the alert with what changed in the last thirty minutes, checks whether the runbook's first three steps apply, and either resolves it, drafts a remediation for a human to approve, or escalates with a summary attached.
Teams describe the human page rate falling by half to two-thirds. The pages that do reach a person arrive with context: what the agent saw, what it tried, what it recommends. Nobody is reading a raw stack trace through one eye anymore.
The job that replaces the pager
The interesting part isn't fewer pages. It's what the engineers are doing instead, because it's not nothing.
- Writing the boundaries. Someone has to decide what the agent may do alone (restart a stateless pod), what it may propose (scale a database), and what it may never touch (anything involving customer data deletion). This is policy work, and it is harder than the on-call work it replaced.
- Reviewing the agent's night. Morning stand-up now includes a read of what the agent did at 3 a.m. The engineers call it "reading the log of the colleague who doesn't sleep."
- Fixing root causes. With the firefighting absorbed, the backlog of "we should really fix that" items is finally getting worked.
"On-call used to be about reaction speed. Now it's about judgment: deciding in advance what we trust the system to do without us. That's a senior engineer's job, and it turns out we needed more of it than we had."
The culture shift nobody planned for
Two second-order effects are showing up. First, junior engineers are losing the traditional way they learned production: by being paged into it. Teams are starting to deliberately route some incidents to humans for training, which would have sounded absurd two years ago. Second, the "hero" culture of the engineer who saves the night is fading, and with it a certain kind of status. Some teams are mourning that more than they expected.
What makes it work
The teams reporting the best results share one trait: the agent reads from the same live event stream the rest of the system does, not from a copy refreshed every fifteen minutes. An agent reasoning about an incident with stale telemetry is worse than no agent, because it is confident and wrong. The pager is becoming an agent's shift only where the agent can see what's happening right now.
Agents are absorbing the reactive half of on-call. The engineers' job is shifting from answering pages to defining what the agent is allowed to do, which is harder, more senior work than the pager ever was.
Signals shortlist
- Teams running agents on operational telemetry report human page rates down 50–66%.
- Pages that reach a human now arrive with agent-generated context: what changed, what was tried, what is recommended.
- New work categories: autonomy policy, morning review of agent actions, and deliberate "training incidents" for junior engineers.
Launch-review draft. Figures in this story are illustrative and drawn from Redpanda field conversations, support data and public sources as noted. They are pending validation by the Redpanda data team before external citation.
- Interviews with platform and SRE teams operating LLM agents on incident data, 2026.
- Redpanda community and developer advocacy conversations.
- Public engineering blogs on agent-assisted incident response.
Ground Truth is written and edited by Redpanda and invited contributors. Research methodology and sources are listed in each story.
