Levi Gans
A page fires at night. The on-call engineer opens five dashboards, checks what shipped today, searches Slack for the last time this happened, and only then starts thinking about the actual cause. That first stretch is the part AI SRE tools try to take off your plate.
An AI SRE is an AI agent that takes an alert or an incident, investigates it across your telemetry, code and recent changes, and reports the likely cause with the evidence behind it. Some also propose a fix, usually a pull request or a command that a person approves. It doesn’t replace your observability stack, your incident process or your engineers. It reads the tools you already run.
The short answer, if you’re comparing the best AI SRE tools for investigating production incidents. Start from where your incidents already live. On Datadog, try Bits Investigation. On incident.io, PagerDuty or Rootly, try their built-in agent. Spread across many tools, or need it in your own cloud, look at Resolve AI, Traversal, Cleric and NeuBird. If your “incidents” are mostly exceptions in a web app, you may not need an AI SRE at all. That case is at the end.
Whatever the vendor, the loop looks roughly like this:
- An alert firesor an incident is declared
- Gather contexttelemetry, deploys, code, past incidents
- Test hypothesesand rule out the ones that don't fit
- Report the causewith evidence and a confidence level
- Propose a fixa pull request, rollback or command
- A person approvesand checks that it worked
AI SRE tools at a glance
| Tool | Lives in | Starts from | Can change things |
|---|---|---|---|
| Datadog Bits Investigation | Datadog | Monitor alert or a prompt | Code fixes, infra actions on approval |
| incident.io Investigations | incident.io | The incident and its alert | Only a PR you merge |
| PagerDuty Paige | PagerDuty | Incident, Slack, escalation policy | Write tools you allow |
| Rootly AI SRE | Rootly | Alert, incident or Auto-run rule | Whatever your credentials allow |
| Resolve AI | Standalone | Pages, as primary on-call | Write tools you allow |
| Traversal | Standalone | Alerts, incidents, Slack, MCP | Read-only unless you grant writes |
| Cleric | Standalone | Alerts and released changes | Fix PRs, read-only on infra |
| NeuBird | Standalone | Alerts from PagerDuty, Slack | Read-only, actions need approval |
| Interfere | Web app error tracker | App errors, not alerts | Your coding agent writes the fix |
Sources are linked in each tool’s section below.
What to check before you buy
Every vendor here says it finds root causes. These questions separate them faster than a demo does.
- Can it change production? Ask which credentials it needs and which tools can write. Rootly’s own docs tell you not to treat its agent as read-only, which is the kind of honesty you want.
- What can it read? Logs, metrics and traces, but also code, deploys, feature flags, runbooks and past incidents. An agent can’t find a cause in data it can’t see.
- Where does it run? Vendor SaaS, a single-tenant instance, your own cloud account, or an agent inside your cluster. This decides how long security review takes.
- Does it say when it doesn’t know? A good report links each claim to a query, a log line or a commit, and marks itself inconclusive when the evidence is thin.
- How is it billed? Most of these meter usage. An agent that runs on every alert can cost more than one you start by hand.
Built into your observability platform
If most of your telemetry already lives in one platform, its own agent has the shortest path to the data.
Datadog Bits Investigation
Bits Investigation is Datadog’s AI SRE agent. You can start it from a monitor alert, a Synthetic test, RUM, a plain prompt, or by replying in Slack. Starting it from Error Tracking is in preview. Or turn on Auto-Investigate so it runs whenever a monitor enters alert.
It reads metrics, APM traces, logs, dashboards, events, Change Tracking, RUM, Error Tracking, Database Monitoring, the Continuous Profiler and your source code, GitHub only. Grafana, Dynatrace, Sentry, Splunk, ServiceNow and Confluence are in preview. Each run ends with an evidence-backed conclusion or is marked inconclusive.
Bits Remediation generates a code fix through Bits Code when the cause is in code. By default you create the pull request yourself. One-click infrastructure actions on Kubernetes and AWS are in preview, and run only when an admin-defined guardrail covers the action and the resource. It’s billed through Datadog’s AI Credits and isn’t supported on Datadog’s US government sites.
Built into your incident tool
These run where your incidents are already declared and paged. They suit teams whose incident process is the center of everything.
incident.io Investigations
Investigations start from the alert that fired and the message that declared the incident. They search Slack, similar past incidents, runbooks, deploys, feature flags, code changes and telemetry in parallel, and check whether a third-party provider was down at the same time. They keep running until the incident is resolved.
Each finding carries a conviction level, from “Speculation” up to “Validated, alternatives ruled out”, and discounted theories stay in the audit trail. Per its trust and safety page, the only change an investigation can make is a pull request a person reviews and merges. By default it opens one only when you ask. With the desktop app, you can pull a live investigation into Claude Code, Codex or Cursor and steer it from there.
PagerDuty Paige
Paige is PagerDuty’s SRE agent. You use it from Slack, the incident details page or the Operations Console. Turn it on for an escalation policy and it investigates the moment an incident triggers, in parallel with your on-call engineer, without delaying the page. Microsoft Teams support is in early access.
It reads logs, metrics and traces through more than 20 connectors, including Datadog, Grafana, New Relic, Splunk and CloudWatch, plus runbooks and code. Tools are split into read-only and write/delete groups, and you set each to “Always Allow” or “Blocked”. It needs PagerDuty AI, and each request to Paige uses four AI Actions.
Rootly AI SRE
Rootly AI SRE starts from an alert or incident, from Slack, or automatically through an investigation rule set to Auto-run. It queries your observability, code, cloud and knowledge providers. Rootly’s account team turns it on for each team.
A finished run says “Root cause identified” with a High, Medium or Low confidence label, or says it’s inconclusive or blocked by missing access. Rootly checks the recorded evidence before it accepts a cause, so the model can’t declare one on its own. Its FAQ is blunt: don’t treat AI SRE as read-only, because some connector tools can change data when the account allows it. Use read-only credentials. Its product page says it “never auto-remediates without human sign-off.”
Standalone AI SRE agents
These connect to whatever you already run, so they suit teams spread across several observability and incident tools.
Resolve AI
Resolve AI can sit first on your escalation policy as primary on-call. It takes the page, decides right away whether to close it or escalate to a person, and runs a deep investigation either way. It escalates whenever it’s unsure, and you can run a trial period where it only records “Would close” or “Would escalate”. It also works in incident channels.
Closing and acknowledging pages happens without approval. Everything else follows tool permissions: each write tool is Blocked, Needs approval or Always allow, and integrations stay read-only until you grant write access. Its GitHub app opens a pull request when you ask for a fix, and never merges. An optional Satellite runs in your cluster, read-only, and keeps credentials behind your firewall. Usage is billed in credits against your contract.
Traversal
Traversal is built for large, complex systems. It triages alerts and investigates incidents, and you can run it from the web, Slack, Teams, an API, or MCP clients like Claude and Cursor. Its integrations cover 19 telemetry tools, including Datadog, Splunk, Prometheus and Sentry, plus GitHub, GitLab and incident tools. It’s read-only by default unless you give it write access.
Its security page offers on-premises deployment, bring your own cloud and bring your own model. With self-healing turned on, it applies a fix, checks it against production and confirms the problem is gone. You choose which kinds of fix run on their own, and Traversal suggests starting with the ones you trust. Getting started means booking a demo.
Cleric
Cleric takes the pager for your alerts and closes false positives, so you’re paged only when you’re needed. It also verifies released changes against production, and keeps watching signals that need more time before it reports.
It reads logs, metrics, Kubernetes, cloud accounts, alerts, code and docs, and works in Slack or Microsoft Teams. Fixes arrive as pull requests for your review. Its security page describes a single-tenant instance per customer, read-only access to infrastructure, and a branch protection rule that keeps it from pushing to your default branch. There’s a free trial.
NeuBird
NeuBird, formerly Hawkeye, investigates alerts from PagerDuty, Slack and other alerting tools. It queries metrics, events, logs and traces where they live, reads AWS, Azure and GCP, and pulls runbooks from Confluence and Google Drive. An MCP server brings it into Claude Code, Cursor and GitHub Copilot.
Its security docs say every connection is read-only, telemetry stays in memory only for the length of an investigation, and you can deploy it as SaaS, with your own LLM, or fully inside your AWS account. Its homepage describes three policy levels. Suggest and Recommend change nothing. Act runs through your own tooling, and every action waits for a person’s approval.
Are AI incident investigation tools worth it for a mid-size team?
Usually, if two things are true. You get paged often, and your incidents cross services, so most of the first stretch goes to gathering context rather than fixing. That gathering is exactly the work these agents automate.
They’re harder to justify when you’re paged a few times a month, or when most incidents are bugs in one app. Setup isn’t free either. You connect each data source, create read-only credentials, and get it through security review.
Then there’s cost. Datadog bills Bits through AI Credits. PagerDuty charges four AI Actions per Paige request. Resolve meters chats, tasks, investigations, incidents and triages as credits. Datadog’s defaults show the risk well: one automatic investigation per monitor every 24 hours, and at most 20 per 15 minutes per organization. Turn on auto-investigate for every monitor and check your bill a week later.
The cheapest test is your own history. Run a trial against five recent incidents you already understand. Did it find the cause? Did it cite real evidence? Did it say “inconclusive” when it should have? Datadog even tracks mean time to conclusion in its Reports tab, so you can measure the time saved instead of guessing.
Where Interfere fits
Interfere is an error tracker for web apps that sorts each problem by how many users it affects, shows which customer accounts and customer tiers it hits, investigates the likely cause, and hands it to your coding agent over MCP.
Tools built for on-call assume incidents are rare and loud. That assumption is breaking. When coding agents push hundreds of pull requests to production a day, bugs get smaller and more frequent, and most of them stay below any alert you’d set. Somebody still has to notice them, decide which ones matter and get them fixed. We built Interfere to do that first pass, so the people on your team can stay sane at that pace.
It isn’t an AI SRE, and it doesn’t try to be one. An AI SRE starts when something pages. Plenty of costly bugs never page anyone. An export that fails for one customer’s admins, or a settings page that throws on Safari, won’t move a dashboard or trip an alert. It shows up as a support ticket a week later.
Those are the bugs we built Interfere to catch. It detects new problems and regressions as production errors arrive and groups related occurrences, so a rare error gets its own row in the Inbox instead of sinking under a noisy one. The Inbox sorts by affected users and shows which customer tiers each problem reaches, from tiers you set through our API. A problem hitting three users on Enterprise is easy to spot and easy to bump up. We think that’s a better signal for an app team than alert volume.
Each problem pulls in traces, logs and session replays. When the investigation can show the cause in code, it attaches a suggested fix, and your coding agent picks it up over MCP. Interfere watches recovery and reopens the problem if it comes back.
It doesn’t take pages or read infrastructure alerts. If your incidents start in Kubernetes or a database, pick one of the tools above. SDKs cover Next.js, React with Vite and TanStack Start, with OpenTelemetry for the rest. No mobile SDKs or self-hosting. See AI root cause analysis tools, or try Interfere.
Which one to pick
For on-call engineers and for MTTR, the biggest win is an agent that starts before a person opens the page. Bits Auto-Investigate, Paige on an escalation policy, Rootly’s Auto-run rules and Resolve as primary on-call all do that.
- Your telemetry is mostly in Datadog: Bits Investigation. It already has the data and the permissions model.
- Your incident process runs through incident.io, PagerDuty or Rootly: their agent already has the incident context. That’s the shortest setup.
- You want an agent to take the first page: Resolve AI or Cleric.
- You need it inside your own cloud: Traversal or NeuBird.
- Your costliest problems are web-app bugs that never page anyone: Interfere. It shows the users, accounts and tiers each one hits, and your own coding agent writes the fix.
For more on cutting resolution time with or without an agent, see how to reduce MTTR.
FAQ
What is an AI SRE?
An AI SRE is an AI agent that investigates production alerts and incidents the way an on-call engineer would. It reads telemetry, deploys, code and past incidents, then reports a likely cause with evidence and a confidence level. Some also propose a fix for a person to approve.
Will an AI SRE replace site reliability engineers?
No. It takes the first pass of an investigation, and engineers still own the decisions. Every tool here lets you limit what it can write, and most put changes behind a person’s approval.
How is an AI SRE different from AI in an error tracker?
An AI SRE starts from an alert or incident and works across your whole stack. An error tracker’s AI starts from one bug and works through your code. If your problems are mostly application bugs, compare Sentry Seer alternatives too.




