TL;DR
Get business pricing on networking and server gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
AI SRE tools are moving from conference demos into company-built systems and commercial incident platforms, amid reports of rising deployments and on-call pages. Their current use is mainly investigation, triage and alert filtering; claims of agents autonomously fixing production incidents remain ahead of what the report found in practice.
AI-assisted site reliability engineering tools are gaining attention as companies report more software changes and on-call pages, but current systems are mainly being used to investigate incidents, triage alerts and filter less important pages—not to make unsupervised production fixes. A report by The Pragmatic Engineer, drawing on conversations with Rootly, incident.io and ResolveAI, describes a field moving beyond conference demonstrations while leaving the case for autonomous remediation unproven.
The report says the recent interest follows a period when AI SRE attracted little visible attention. It points to OpenAI’s Sevbot, an internal AI SRE bot, an engineer at ING describing work on an in-house tool, and a growing presence of vendors at technology conferences. These examples suggest experimentation, but do not establish how widely the tools are deployed or how well they perform across companies.
The main use described is accelerating incident investigation, often inside Slack or Microsoft Teams. Systems can gather information from company tools and records, propose possible causes, and help engineers steer an investigation. Vendors and in-house teams also use a “context layer”—including a knowledge graph or stored operational memory—to give agents relevant information about a particular system.
Some systems split work among multiple agents, while teams test proposals with different models or adversarial checks. The report says companies also use evaluations and backtesting, but measuring accuracy is difficult: the right answer depends on a company’s infrastructure and on the engineering team’s own standards. Speed, cost and accuracy can pull in different directions, prompting teams to use model routing, shorter prompts and cache monitoring.
More Pages, Different On-Call Work
The practical appeal is a chance to reduce time spent assembling evidence during an incident and, eventually, to lower the number of pages that reach engineers. The report cites Rootly data showing pages have risen from a year earlier and out-of-business-hours pages have nearly doubled. It links the broader pressure to higher software change rates, which the report says have at least doubled at many companies; these are reported trends, not a universal measurement of the industry.
That shift could change the work rather than remove it. If agents handle routine triage, engineers may receive fewer alerts but face more difficult incidents when they are called. There is also a deskilling risk if engineers lose practice diagnosing failures. The report suggests incident-response simulations, including training offered by Uptime Labs, as one way to maintain those skills.
AI-generated code can also introduce new failure patterns. The report describes an emerging category of AI-related outages, though it provides no overall count or measured rate. For operations teams, the value of AI SRE will depend not just on whether an agent can produce a plausible explanation, but on whether it can do so reliably within the safeguards and systems of a specific organization.
Products Worth Considering
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
As an affiliate, we earn on qualifying purchases.
From Azure Demo to Incident Agents
At Microsoft’s May 2025 BUILD conference, the company demonstrated an Azure SRE Agent responding to an alert, identifying a root cause and shipping a fix without waking a human on-call engineer. The report’s author questioned whether teams would let an AI make production decisions without oversight, given the possibility of worsening an outage. The demonstration illustrated a goal, not evidence that such autonomous operation is routine.
The report says the term “AI SRE” has stuck despite being imprecise: site reliability engineers do much more than respond to alerts, and vendors themselves have not settled on a better label. Proposed alternatives include automated incident investigation or remediation agents. The author disclosed an investment in incident.io and said they had not been paid by the companies discussed.
One simple reliability model in the report describes incident rate as the number of changes multiplied by the probability that a change causes failure. More deployments can raise exposure if the risk per change does not fall. AI SRE tools principally aim to help teams learn from incidents and improve response; testing and verification before deployment are a separate way to reduce failure risk.
“Incident rate = [Change rate] x [Probability]”
— The Pragmatic Engineer report
How Reliable Are These Agents?
The report does not establish a common accuracy rate, cost benchmark or independent comparison of AI SRE products. Evaluation is company- and team-specific, and vendor examples do not by themselves show that tools reduce outage duration or prevent failures. The cited Rootly page data also comes from one provider; the report does not give enough information here to assess its sample, exact time window or comparison method.
It remains unclear how often companies allow agents to make or deploy production changes, what safeguards govern those actions, and how much human review is required. The report characterizes alert triage and suppressing unimportant pages as more established than autonomous fixes, but does not quantify adoption. The reliability and consequences of AI-related outages are also not measured.
Evidence Will Shape Wider Adoption
The next test for the category is whether organizations can show, with their own incident records, that agents improve investigation speed or reduce unnecessary pages without creating new operational risks. Teams will need to define what counts as a correct diagnosis, test systems against past incidents, and decide where human approval is required—especially before any production change.
For now, the report points to continued experimentation by vendors and in-house engineering teams rather than a settled move to autonomous on-call. Further adoption will depend on performance in real incidents, the cost and latency of running agents, and whether engineers can retain the skills to take over when an automated investigation is wrong or incomplete.
Key Questions
What does AI SRE mean?
It is a broad label for AI-assisted on-call and incident-response tools that investigate alerts, gather operational context, help identify possible causes or triage pages. The report says the term is imprecise because SRE work extends well beyond on-call response.
Can AI SRE tools fix production incidents on their own?
The report describes autonomous fixes mainly as a conference-demo ambition. Current uses it discusses focus more on investigation and triage; it does not establish that unsupervised production remediation is common or dependable.
Why are companies exploring these tools now?
The report links interest to rising deployment activity and on-call pages. It cites Rootly data showing pages increased year over year and out-of-business-hours pages nearly doubled, while warning that these figures do not establish a universal industry trend.
How do teams check whether an AI SRE agent is accurate?
Approaches described include evaluations, backtesting, multiple models and adversarial review. There is no shared accuracy benchmark in the report; what counts as a useful result varies by company and engineering team.
Could AI SRE change engineers’ skills?
Potentially. If routine alerts are filtered or investigated automatically, engineers may receive fewer pages but have less practice with incident response. The report flags deskilling as a risk and points to simulations as one way to keep response skills current.
Source: newsletter.pragmaticengineer.com
