With AI agents and AI-generated code becoming the norm in modern enterprise software, backend systems are evolving faster than ever. And it’s leaving most operations teams with an impossible choice: burn out senior talent on repetitive firefighting, or hand production over to untrained AI agents.

With disruptions costing enterprises an average of $300,000 per hour, manual firefighting isn’t an option. And beyond the dollar costs, each incident that drags half your team into a triage session grinds your product roadmap to a halt. 

Right now, AI looks like the answer—59% of organizations are already turning to AI agents to increase speed and capacity. But throwing nondeterministic AI at an operations system can be just as costly as the manual approach. Without a strong foundation and continuous learning loop, AI is ineffective at best. At worst, it executes poor decisions faster than your SREs can react.

At PagerDuty on Tour 2026, David Williams, Senior Vice President of Product at PagerDuty, laid out our vision for autonomous operations: a world where an AI agent reliably resolves incidents before it ever pages your team at 3 a.m.

Scaling engineering capacity today without adding systemic risk requires a self-improving AI operations system that continuously captures live human triage and turns it into compounding operational intelligence. Already, nearly half (48%) of organizations have improved resilience by turning incidents into structured learning.

Here’s how teams can build a self-improving system that learns from human workflows to create lasting resilience in five steps:

1. Build the foundation on system telemetry and service ownership

Before an AI system can learn, it needs a unified foundation of system data and mapped service ownership to make sense of your environment.

The first step is to route disparate monitoring data from multi-cloud infrastructure, APMs, and CI/CD pipelines into a single platform that normalizes it into a clean, machine-readable format. According to PagerDuty’s 2026 State of AI-First Operations Report, more than half (51%) of organizations that consolidated fragmented tools into a single platform saw improved operational resilience compared to those relying on other strategies.

But telemetry is only part of the picture. During an incident, it just tells your team what is failing. It doesn’t answer who is responsible. Without explicit service mapping, a high-severity alert just gets passed back and forth until someone claims it, while MTTR grows.

Dave Hogan, Principal Architect of Service Reliability Engineering at Workday, made a similar point at PagerDuty on Tour 2026 in London: “If [teams] do something really simple, like just assigning accountability and ownership to their products and technical services, what they’re gonna see is a pretty dramatic uptick in their resilience.”

By structuring your raw telemetry, service relationships, and team ownership into a single connected fabric, PagerDuty gives AI agents a clean map to navigate your operations. That map forms the foundation for the AI to learn and improve.

2. Capture live operational data as it happens

Without human context, standalone AI models suffer from model collapse. Trained on their own outputs instead of fresh data, they drift from reality and produce flawed responses.

New, human-generated data is essential for sustaining the quality of agentic workflows over time. Agents evolve fastest by learning directly from your engineers during active incident response. PagerDuty’s built-in AI Agents turn real-time human triage into compounding intelligence in three steps:

1. PagerDuty connects to your monitoring stack and communication platforms under strict, deterministic guardrails you set.
2. Plugged into escalation policies, the SRE Agent becomes your team’s first line of defense. It receives alerts first, correlates signals, checks recent deployments, and assembles a diagnostic summary before a human is ever paged. You control which diagnostic actions it’s allowed to take.
3. Working natively in Slack alongside the team, the agents track hypotheses, log tested solutions, and turn the conversation into structured operational intelligence.

Before the agent is allowed to execute any runbooks autonomously, it focuses on assisting with triage, while capturing and converting human knowledge into permanent operational data.

3. Auto-generate post-incident reviews

PagerDuty’s 2026 State of AI-First Operations report found that AI-first organizations expect a massive need for stronger post-incident learning over the next year.

But simply filing a post-incident review (PIR) isn’t enough if it doesn’t turn into action. After a service disruption, critical context often vanishes into message archives or untouched databases—meaning it never actually improves the system going forward.

According to Opsera’s benchmark data on incident management, teams that complete their PIR action items keep repeat incident rates below 10%. But teams that write PIRs and don’t execute on them see repeat rates of 20% to 30%. At that rate, they’re just producing more paperwork, not operational learning.

PagerDuty automatically synthesizes incident timelines, discussions, and hypotheses into a structured post-incident review. It turns the PIR into checklists for your team to execute and feeds the knowledge to agents that use it to improve their future actions. The system itself works better based on what it’s learned, and eliminating the toll of manual reporting shifts engineering hours from documentation to deeper root cause analysis.

4. Delegate autonomous execution

With deep operational memory established, the SRE Agent can now safely move from passive alert routing to controlled autonomous execution. Working within strict boundaries you set, agents execute diagnostic routines, apply approved runbooks, and resolve known failure modes without pulling a human into the loop.

Consider a midnight scenario where web servers suddenly start timing out. PagerDuty’s SRE Agent receives the alert first, queries system logs, and correlates the pattern against 11 historical incidents in its operational memory. Once it recognizes the failure mode, it executes an approved runbook to scale server capacity, verifies recovery, closes the ticket, and publishes a completed PIR. It resolves the disruption before an on-call engineer is ever paged.

When an incident falls outside the agent’s well-understood framework, the agent escalates to responders with full diagnostic context and tested hypotheses so they can engage right away.

5. Shift left to prevent the next incident

The first four stages create a highly efficient incident response machine, which dramatically reduces MTTR. But the best incident is one that never happens at all.

Applying critical checks earlier in development turns accumulated operational memory into a shield. PagerDuty shifts service health history, dependencies, and incident data left, directly into everyday developer tools.

Model Context Protocol (MCP) integrations connect PagerDuty with developer environments like GitHub Copilot, Cursor, and Claude. Developers and their coding agents can query your organization’s historical operations data in natural language before new code ever reaches deployment.

For example, an engineer preparing to deploy code to a core payment service can ask GitHub Copilot to check system stability first. PagerDuty’s agent reviews the service history and flags a problem: The target service is healthy, but four critical downstream dependencies are experiencing severe disruptions and unresolved backlogs. So the developer pauses the release, coordinates with those teams, and averts a high-severity incident before pushing a single line of code.

Operational intelligence is the new competitive moat

As AI-assisted development increases service complexity, operational resilience is becoming a core business differentiator. Knowing your system is secure lets your team ship faster, protect margins, and guarantee ambitious SLAs.

But navigating that growing complexity requires more than raw automation or a bigger engineering team. To build resilience in the agentic era, you need an agent that knows what your engineers already carry inside their heads.

PagerDuty’s SRE Agent takes individual knowledge and experience and turns it into shared operational memory. When human problem-solving continually feeds autonomous agents, disruptions become less frequent and easier to resolve when they do happen.

See how PagerDuty turns operational intelligence into a self-improving system for your engineering team. Schedule a demo today.