In May 2025, on Microsoft’s BUILD stage, an alert fired inside a demo system. The Azure SRE Agent found the root cause and shipped a fix on its own. No engineer woke up for any of it. The demo was clean, fast and complete. What production teams actually run, eighteen months later, is narrower and more interesting.
The direct answer: an AI SRE is an agent that joins an incident and investigates it. It reads the alert, the logs, the recent deploys and the runbooks. It posts findings where the team already works. At the careful shops, it proposes fixes and then waits. Engineers still decide. Engineers still carry the pager. The investigation got faster. The authority did not move.
Investigation is the product being sold
The work happens inside Slack. An incident channel opens. The agent fans out sub-agents, one per hypothesis. Engineers steer it with plain instructions in the channel. Ask it to check a system. Ask it to pull specific logs. The summary waits when the on-call engineer joins, about five minutes after the alert.
Every serious tool sits on a context layer. The layer maps services, data stores, cloud resources, deployments, runbooks and past incidents. Rootly calls its layer Atlas. incident.io calls its layer Nexus. ResolveAI calls its layer HiveMind. The names differ. The contents barely do.
The vendors split on the fix step. incident.io passes incident context to the engineers’ own coding agents. ResolveAI writes simple fixes itself and hands larger changes over. Rootly aims at agents that draft a fix and merge it when a second agent judges it safe. That last step is still a goal. Today, its merges wait for a human approval, like everyone else’s.
One vendor designed its agent to stay quiet by default. It posts only when a question in the channel sits unanswered. Or when a stakeholder joins without context. Or when someone declares an incident mitigated without proof. Or when a risky action appears with no rollback plan. Silence, it turns out, is also a feature that teams grade.
Why the pager got louder first
AI coding agents roughly doubled how often many teams ship code within a year. More changes per week means more chances to break production per week. The incident rate follows the change rate. That is the demand under this whole category.
Rootly’s own incident data shows pages up on last year. Pages outside business hours have nearly doubled. Treat that as vendor data. No public dataset sits behind it, and the direction matters more than the precision. The buyer’s question is not whether the agent is clever. The question is whether your on-call rota survives your own deploy rate. Meta ran internal systems like these two years ago. The pattern has barely moved since. What changed is that the pattern now has a price list.
The operability test is the same one that decides agent frameworks. Six frameworks want to run your agents. Pick the one your team can run half-asleep asks who can read a failure at 2 a.m. An investigator you cannot inspect mid-incident fails that test in the same way.
The verifier is the part to copy
One vendor rebuilt its investigation agent three times in two years. The scripted version handled common incidents and failed on odd ones. The single-agent version graded its own hypotheses. It nearly always confirmed its first suspicion. Self-grading failed in production, in the open.
The current design splits the jobs into three roles. One agent investigates. A second, separate agent challenges the findings. A third keeps stakeholders updated. The challenger exists because the earlier design could not doubt itself. That separation is the most useful idea in the category, and it costs no licence to copy.
Large teams reached the same shape in-house. OpenAI built an internal bot that gathers details and proposes mitigations. It never runs a mitigation on its own. An engineer instructs it first, every time. A large European bank is building its own version for the same reason. These tools need broad access to telemetry, logs and code. The larger the company, the less of that access it hands a vendor. Stripe’s in-house agent averages about three minutes on a simple investigation. A deep one averages fourteen.
There is a cost side to log. The incidents that still reach engineers will be the hard ones. A junior engineer who never runs a manual investigation loses the skill the agent imitates. The teams testing these tools pair them with incident simulations for exactly that reason.
The signature rule already exists for documents. Review AI-generated work with one signed line before your name goes on it sets that bar for AI drafts. Incident findings deserve the same rule. An agent’s root cause is a draft until an engineer signs it. Who signs an AI agent refund after Ada joins Dialpad? asks the same question about money. On-call is where the question costs sleep instead.
The runs are not free, either. Stop counting tokens. Even the man who sells them calls the metric fake covers why the token bill is the wrong scoreboard. The right scoreboard is engineer minutes returned per incident.
Put one question to any AI SRE vendor, in writing, before a pilot. Show us the last incident where your agent reached the wrong cause. Show us what stopped it acting on that answer. A vendor with a verifier story answers fast. A vendor with only a stage demo does not. No hands-on test of any AI SRE tool ran for this piece.
