AI SRE Agents in 2026: What They Can Really Do

An AI SRE agent is an autonomous AI agent that triages alerts, investigates incidents, finds the root cause, and proposes or executes a fix — without step-by-step human direction. In 2026 they moved from demo to product: on July 27, Dynatrace launched autonomous SRE agents, joining Azure, Datadog, and New Relic. But the honest headline is narrower than the marketing: agents now investigate on their own, while a human still approves the risky fix.

TL;DR

The AI SRE agent — an autonomous teammate that works your on-call queue — became a real product category in 2026, not a research demo. Dynatrace's July 27 launch added an Autonomous SRE Agent, a multi-cloud Cloud SRE Agent, and a no-code Agent Builder to its platform.¹ It follows Microsoft's Azure SRE Agent (generally available since March)², Datadog's Bits AI SRE, and New Relic's SRE Agent.³⁴ These agents genuinely do the tedious part of incident response — parallel triage, cross-signal root-cause analysis, postmortem drafts — and vendors report large drops in mean time to resolution. What they mostly do not do yet is push high-risk changes to production unattended: rollbacks, scaling, and restarts still route through human-in-the-loop approval. Gartner expects agentic AI to run most enterprise IT infrastructure by 2029, but also warns that more than 40% of agentic AI projects will be scrapped by 2027.⁵⁶ Both things are true at once.

What you'll learn

• What an AI SRE agent is, and how it differs from traditional AIOps
• What Dynatrace actually launched on July 27, 2026
• Which vendors ship AI SRE agents today — Azure, Datadog, New Relic, and more
• How an AI SRE agent works through a live incident, step by step
• Why autonomous remediation still needs a human to approve the fix
• What the skeptics — and Gartner — say about the category
• Whether AI SRE agents replace on-call engineers

What an AI SRE agent actually is

An AI SRE agent is software that does the job of a site reliability engineer during an incident: it reads the page, gathers evidence across your telemetry, forms a root-cause hypothesis, and either recommends or carries out a remediation. The word that matters is agent — unlike a fixed automation script, it decides what to investigate next based on what it finds, the way a human on-call engineer would.

This is the distinction most vendor pages blur: an AI SRE agent is not the same thing as traditional AIOps. Classic AIOps is about reducing noise — grouping related alerts, suppressing duplicates, and correlating anomalies so a human gets one page instead of fifty. That is valuable, but it stops at "here is the problem." An AI SRE agent picks up where AIOps leaves off and does the investigation: it correlates logs, metrics, traces, and topology in parallel, ties a latency spike to the deploy that changed a database connection-pool setting, and writes up what it found.⁷ AIOps clusters the signal; the agent chases it down.

A useful way to keep the categories straight: AIOps tells you the house alarm is going off, an AI SRE agent walks the rooms and tells you which window is broken. This is also different from observing your own agents — instrumenting an LLM app with traces so you can debug it — which is its own discipline. (If that is what you are after, see our guide to tracing agent tool calls with OpenTelemetry; it is a separate problem from agents that operate your infrastructure.)

What Dynatrace launched on July 27

On July 27, 2026, Dynatrace announced a set of additions to its Dynatrace Intelligence platform that push it "beyond analysis and recommendations into trusted, autonomous execution."¹ Three pieces matter for on-call teams.

The Autonomous SRE Agent triggers on its own when a new problem is detected, decides whether that problem belongs to an investigation already in flight, and enriches the ongoing investigation with additional context instead of spawning a duplicate. The Cloud SRE Agent coordinates remediation across AWS, Microsoft Azure, and Google Cloud, and keeps a single auditable record of what happened. And Agent Builder lets customers assemble their own no-code agents for workflows specific to their environment.¹ Dynatrace also expanded integrations into ServiceNow, Atlassian, and PagerDuty so the agent works inside the tools teams already page from.¹

On timing: the Cloud SRE Agent and the enhanced natural-language assistant are available now to SaaS customers, while the Autonomous SRE Agent and Agent Builder are expected in August.¹ Dynatrace's pitch leans hard on trust — it grounds every action in what it calls a "deterministic, real-time" understanding of the environment rather than a probabilistic guess, and wraps it in logging, auditability, and human approval. "Most observability platforms stop at data, leaving humans to find answers, determine what to do, and execute," said chief product officer Steve Tack.¹ The product's job is to close that last gap — carefully.

Who else ships AI SRE agents in 2026

Dynatrace is a late-ish entrant into a field that filled up fast this year. One of the most production-tested examples is Microsoft's Azure SRE Agent, which reached general availability in March 2026.² Microsoft says that across its own services the agent has mitigated more than 35,000 incidents and now saves over 20,000 engineering hours a month — impressive numbers, but self-reported and measured on Microsoft's own infrastructure, not an independent benchmark.²

Datadog's Bits AI SRE has been generally available since December 2025, after debuting in limited availability at the company's DASH conference earlier in the year; Datadog calls it the first generally available agent in its Bits AI lineup.³ New Relic announced its SRE Agent in February 2026 — an "always-on" teammate that diagnoses incidents and recommends next steps, sometimes before an engineer even acknowledges the page; New Relic's own 2026 AI Impact Report claims teams using its AI features resolved incidents 25% faster.⁴ Beyond the observability incumbents, a wave of specialists now compete on AI-driven incident response, including incident.io, PagerDuty, Rootly, Resolve AI, Cleric, and Traversal.

The interesting split is philosophical. Some vendors, like Metoro, position their agent as running the full loop — detect, investigate, verify, and open a fix. Others, like Honeycomb, deliberately keep a human in the driver's seat and cap the agent at investigation and recommendation. That divide — how much you let the agent do versus merely say — is the real decision buyers face in 2026, and it is more important than any leaderboard.

How an AI SRE agent works through an incident

Walk through a 2 a.m. page and the loop looks like this. The agent starts the moment an alert fires, not when a human logs in. It pulls the relevant telemetry — logs, metrics, traces, and the service topology — and correlates them in parallel rather than in the linear, one-query-at-a-time way a sleepy human would. Instead of reporting "latency is high," it produces a causal statement: latency climbed right after a deploy that changed a connection-pool setting, and here are the three services downstream that are now affected.⁷

From that hypothesis it drafts the next steps: the probable root cause, the supporting evidence, and a proposed remediation — roll back the deploy, scale the fleet, restart the unhealthy pods. For low-risk, reversible actions inside its approved scope, a well-configured agent may execute directly and then watch to confirm the metrics recover. For anything with real blast radius, it stops and asks: it posts the proposed fix with evidence into Slack or your incident tool, a human approves, and only then does it act. Afterward, it drafts the postmortem while the timeline is still fresh. The pattern is the same across vendors — investigate autonomously, remediate under supervision.

Why humans still approve the fix

Here is the part the launch posts underplay. In 2026, autonomous investigation is real and reliable; autonomous remediation — an agent making unattended, high-impact production changes — is mostly still supervised. The reasons are practical: a wrong rollback or an over-aggressive scale-down can cause a bigger outage than the one being fixed, and a probabilistic system occasionally guesses wrong. So the shipping products gate risky actions behind human approval by design. Dynatrace frames its entire release around "human oversight and governance," and its Autonomous SRE Agent's actual job at launch is enriching investigations, not pushing changes to production on its own.¹

This is not a knock on the technology — it is the correct engineering posture, and it is what separates a durable product from a demo. Practitioners who have run these agents in anger tend to describe fully hands-off remediation as a next-year capability, not a today one: expect supervised mode, with the agent proposing and a human approving high-blast-radius actions, for a while yet. The honest 2026 value proposition is not "fire your on-call rotation." It is "your on-call engineer stops doing the first 30 minutes of grunt-work triage on every page." That is a big deal on its own.

What the skeptics — and Gartner — say

The analyst view captures the tension neatly. Gartner's Predicts 2026 report expects that by 2029, 70% of enterprises will use agentic AI to help operate their IT infrastructure, up from less than 5% in 2025 — a steep adoption curve that squares with how much vendors shipped this year.⁵ But the same firm is openly skeptical of the current hype. In a June 2025 forecast, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, blaming escalating costs, unclear business value, and inadequate risk controls.⁶

Gartner also coined a term worth keeping in mind while reading any AI SRE pitch: "agent washing," the rebranding of older automation, RPA, and chatbots as "agents" without real agentic capability. The firm estimated that only around 130 of the thousands of self-described agentic AI vendors were building something genuinely new.⁶ For a buyer, the takeaway is not cynicism but diligence: ask what the agent decides autonomously versus what is a fixed script, insist on an auditable record of every action, and test against your own incidents before trusting any headline MTTR number. Vendor claims of large mean-time-to-resolution reductions are plausible for triage, but they are marketing figures, not neutral benchmarks, until you reproduce them on your telemetry.

Do AI SRE agents replace on-call engineers?

Not in 2026, and probably not in the way the phrase implies. What they replace is the toil: the repetitive first pass of correlating dashboards, the manual root-cause hunt, the blank-page postmortem. Judgment calls — whether to accept a fix with real customer impact, how to weigh a risky remediation against the outage it prevents, when to declare a major incident — stay with humans, both because the technology is not reliably safe enough for unattended high-stakes action and because accountability still needs an owner.

The likelier near-term outcome is a smaller on-call rotation covering more services, with each engineer supervising an agent that handles the mechanical work. That also reshapes the job: the skill that matters shifts from "who can grep logs fastest at 2 a.m." toward "who can scope an agent's permissions, review its reasoning, and catch when it is confidently wrong." Keeping a human able to stop an agent mid-action is its own emerging discipline — see our piece on AI agent kill switches — and it matters more, not less, as agents get more autonomous. The foundations of the role do not change; if anything, solid SRE practices become the thing that lets you supervise an agent safely.

The bottom line

AI SRE agents crossed from concept to product category in 2026, and Dynatrace's July 27 launch is a clean marker of how mainstream the idea has become among the observability incumbents.¹ The capability is real: these agents genuinely compress the investigation phase of incident response, and the biggest vendors have the production numbers — self-reported as they are — to show for it. The nuance that separates a good buying decision from a disappointed one is autonomy scope. In 2026, the safe, honest framing is that AI SRE agents investigate for you and remediate with you. Buy them to delete toil and shrink mean time to resolution, scope their permissions tightly, keep a human on the approval button for anything with blast radius, and treat every "up to 70% faster" claim as something to verify on your own incidents — not a promise.