Skip to content
← all posts
Agentic SREAI OperationsEngineering LeadershipProduction SystemsAutomation

Agentic SRE: When Your On-Call Engineer Is an AI Agent

At 3:14 AM on a Saturday, an alert fires. A core service is experiencing elevated latency and a 4% error rate.

The traditional response: PagerDuty wakes up a human. They SSH in, bleary-eyed, query Datadog, read logs, check recent deployments, and try to piece together the puzzle while the SLA burns.

The new response: An SRE agent that was already watching the metrics sees the anomaly. Within seconds, it correlates the latency across three dependent services, identifies a runaway query introduced in a deployment two hours ago, and proposes a rollback. All before the human's phone rings.

This is Agentic SRE.


From AIOps to Agentic SRE

AIOps has been around for years, but it was fundamentally passive. It offered better alert routing, smarter log correlation, and noise reduction. It told the human where to look.

Agentic SRE is active. It performs autonomous investigation, diagnosis, and, crucially, remediation.

Passive AIOps

Alert: 'High CPU on Web-04. Correlated with 500 errors.' (Human must fix it).

Active Agentic SRE

Action: 'High CPU on Web-04. Root cause isolated to bad regex. Reverted config. Service restored.'
Early adopters of Agentic SRE at scale are reporting a 40% to 70% reduction in Mean Time To Resolution (MTTR) for routine infrastructure incidents.

The Architecture of an SRE Agent

An SRE agent is a multi-modal system equipped with specific operational tools.

Capability TierDescription
Monitoring IntegrationReads metrics, logs, and distributed traces via APIs (Datadog, Prometheus).
Investigation ToolsQueries dashboards, reads internal runbooks, and checks the Git commit history.
Remediation ToolsExecutes scripts to restart services, scale capacity, or rollback deployments.

The Human-Agent Handoff

The most important design decision in Agentic SRE is the human-agent handoff. When does the agent act autonomously, and when does it escalate?

We implement this via privilege rings:

  • Tier 1 (Autonomous): Low-risk actions. Restarting a stateless worker pod, scaling up read replicas. The agent executes and notifies.
  • Tier 2 (Propose + Await Approval): Medium-risk actions. Rolling back a deployment, pushing a config change. The agent prepares the PR or command, pages the human, and waits for a "yes/no".
  • Tier 3 (Escalate Immediately): High-risk incidents. Data loss risks, security breaches, database schema corruptions. The agent gathers context but takes zero modifying actions.
The goal is not to remove humans from the loop entirely, but to move them to the top of the loop — approving well-researched plans rather than digging through logs at 3 AM.

Safety Rails for Autonomous Remediation

An agent that has the access to fix your infrastructure also has the access to destroy it.

You must implement blast radius limits. An agent might be allowed to restart 10% of the fleet, but a circuit breaker trips if it attempts to restart 50%. You need remediation budgets — if an agent has executed three automated fixes in an hour and the alerts are still firing, it must stop and escalate.

Agents that can fix things can also break things, much faster than a human can. Never grant an SRE agent write access to production without hardcoded, deterministic circuit breakers.

The Evolving Role of the Human SRE

Agentic SRE doesn't eliminate the Site Reliability Engineer; it elevates them.

The human shifts from firefighter to architect. SREs spend their time designing the agent's runbooks, reviewing its past actions to improve its prompts, calibrating its confidence thresholds, and handling the complex, novel failures that the agent can't understand.

The future SRE is the architect of autonomous systems. You are no longer writing bash scripts to fix servers; you are engineering the constraints for an AI that fixes servers.

What Can Go Wrong

The failure modes are new and terrifying:

  • Misdiagnosis: The agent confidently misinterprets a spike in legitimate traffic as a DDoS attack and aggressively rate-limits real customers.
  • The Remediation Loop: The agent restarts a service, the service takes 2 minutes to boot, the agent assumes the restart failed and restarts it again, trapping the system in a permanent boot loop.
Deterministic safeguards must always override non-deterministic agents. If the agent wants to scale down the database, a hardcoded rule must say "never scale below 3 nodes" regardless of the agent's reasoning.

// key takeaway

Agentic SRE is not about replacing humans. It's about letting the agent handle the 3 AM alerts that require routine investigation and known remediation, so humans are reserved for the problems that actually require judgment, creativity, and experience.