TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
AI SRE tools are drawing interest as software teams report more deployments and on-call pages, with vendors offering agents that investigate incidents through tools such as Slack. The reported evidence points to assistance with triage and investigation, not the fully autonomous production fixes shown in conference demos; accuracy, safety and measurable impact remain company-specific and unsettled.
AI tools for site reliability engineering are gaining traction among vendors and engineering teams as companies contend with more deployments and on-call pages, according to a report by The Pragmatic Engineer. The systems in use are mainly designed to investigate incidents and help triage alerts; the report says agents that independently diagnose an outage and ship a production fix without human oversight remain a conference-stage vision, not an established practice.
The report traces the current interest to rising software change rates, which it links in part to AI coding agents producing more code and deployments. It cites Rootly data showing that pages have increased from a year earlier and that outside-business-hours pages have nearly doubled. The source does not provide the underlying counts or a broader industry dataset, so the figures should be read as Rootly’s reported observations rather than universal measures.
The common product pattern is an investigation agent that gathers incident information and works through services such as Slack or Microsoft Teams, with engineers steering its investigation. Vendors Rootly, incident.io and ResolveAI told the report’s authors about approaches that include coordinating multiple agents and giving them a company-specific context layer, such as a knowledge graph and stored operational history. Some companies, including OpenAI and ING, were also reported to be developing internal tools.
Engineering teams are still debating whether these systems should make code changes. The report describes accuracy checks such as testing incident hypotheses with separate agents or different models, and evaluating agents against past incidents. But it says there is no single reliable measure of accuracy: useful performance depends on each company’s systems and even on individual teams. Cost, speed and quality also involve tradeoffs, which vendors address through techniques including model routing and shorter prompts.
More Pages, More Triage Work
If deployment frequency rises, teams may face more opportunities for changes to cause failures. The report frames incident frequency as a product of how often changes are made and the probability that any one change causes a problem. That relationship helps explain the commercial interest in tools that can shorten investigation and reduce the burden on on-call engineers, although it does not establish that AI itself is causing a particular incident increase.
For engineers, the nearer-term prospect is fewer low-value alerts and faster initial investigation, not the removal of human responsibility. The report says AI-assisted alert triage, including silencing unimportant pages, is becoming more common. If routine pages are filtered out, however, the remaining incidents may be harder to resolve. There is also a deskilling concern: engineers who rely heavily on automated diagnosis could get less practice handling incidents. The report points to incident-response simulations as one way teams can maintain those skills.
There is a further operational risk in allowing an agent to change production systems. A mistaken diagnosis or unsafe fix could worsen an outage, so the boundary between recommending an action and executing it is consequential. The source describes that boundary as disputed, not as settled by a proven industry standard.
Products Worth Considering
DJI Osmo Pocket 3 Filter Bundle – Vlogging Camera with 1 CMOS Sensor and 4K/120fps
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Jackery Solar Generator 1000 v2 with 200W Solar Panel
As an affiliate, we earn on qualifying purchases.
From Demo to Incident Investigation
At Microsoft’s BUILD conference in May 2025, the company demonstrated Azure SRE Agent responding to an alert, identifying a possible cause and shipping a fix without waking a human on-call engineer. The report’s author said the demonstration felt artificial and questioned whether teams would permit an AI to make production decisions without oversight. The report does not establish that Microsoft deployed the demo capability in production.
After a period with little public discussion, the author says AI SRE became more visible through vendor conference booths and conversations with engineering teams. The report cites OpenAI’s Sevbot and an in-house effort at ING as examples, while Rootly, incident.io and ResolveAI provided input on how such products are being built. The author disclosed being an investor in incident.io and said they had not been paid to mention the companies.
“AI SRE” has become a common label for this class of on-call tooling, though the report says vendors themselves find it imprecise: site reliability engineers do much more than handle alerts. Other labels, including automated incident investigation agents, have not gained the same traction.
““The fully autonomous agents demoed on conference stages are still not a reality.””
— The Pragmatic Engineer report
Safety and Performance Still Vary
The report does not provide independent, cross-company measurements showing that AI SRE tools reduce outages, shorten recovery times or lower on-call workload. The reported increase in pages comes from Rootly, and its underlying counts and methodology are not included in the supplied material. It is also unclear how widely the named internal tools are used, what tasks they perform in production, or how their results compare with human-led incident response.
There is no shared accuracy benchmark, and success on one company’s infrastructure may not transfer to another’s. The degree of human review, permissions granted to agents, and safeguards against harmful changes also differ or are not detailed. The report’s findings describe an evolving market rather than evidence that autonomous remediation is broadly safe or effective.
Watch for Evidence Beyond Demos
The next meaningful test will be whether teams can show that AI-assisted investigation improves incident outcomes under real operating conditions. Useful evidence would include clearly defined measures, such as time to identify a cause, false or unnecessary pages, recovery time and the rate of incorrect recommendations, alongside information about human approval and safeguards.
For now, the report describes a gradual path: agents help collect context and triage alerts while engineers retain oversight, with teams testing systems against prior incidents and tuning them to local needs. Whether that work leads to more autonomous fixes remains open. On-call engineers are likely to remain responsible for difficult incidents while organizations determine where automation is dependable.
Key Questions
What does “AI SRE” mean?
It is a commonly used label for AI-assisted on-call and incident-response tools. In the report, the main focus is investigating incidents and triaging alerts, rather than covering the full range of site reliability engineering work.
Are AI agents fixing production outages on their own?
The report says fully autonomous agents that diagnose an incident and deploy a fix without human oversight remain a demo-stage idea, not an established industry practice. Current tools are described mainly as supporting investigation and alert triage.
Why are companies exploring these tools now?
The report links interest to more frequent code changes and rising on-call pages. It cites Rootly’s data showing pages were up year over year and outside-business-hours pages had nearly doubled, but gives no raw counts or industry-wide comparison.
Do AI SRE tools work?
The source describes active development and practical uses in investigation and triage, but does not provide independent evidence that the tools consistently reduce outages or recovery time. Accuracy and usefulness vary by company and team.
What risks should on-call teams watch?
Teams must account for incorrect diagnoses or changes that could worsen an outage, as well as the possibility that reliance on automation reduces engineers’ incident-response practice. The report describes human oversight and simulation training as relevant safeguards.
Source: newsletter.pragmaticengineer.com
