Agentic AI · Day-2 Operations

Shipping the agent
was the easy part.

Agent operations include post-launch activities such as SLO monitoring, on-call rotations, incident response, runbooks, and cost dashboards to ensure agents run smoothly at all times, even in the early hours of the morning.

overview SLIDE 01 : AGENT OPERATIONS Cover slide: Agent Operations
Quick answer

Agent operations (AgentOps) Applying SRE principles to maintaining production AI agents involves setting service-level objectives, implementing on-call rotations, responding to incidents, creating runbooks, and managing costs and capacity for systems that have different failure patterns than traditional software. Anomaly detection is crucial due to the variability in agent outputs. On-call procedures are evolving with a tiered autonomy model allowing agents to handle routine incidents independently, while complex judgment calls are escalated to humans. The goal is for agents to handle the majority of on-call responsibilities, freeing up humans to focus on the critical twenty percent that requires human judgment.

◷ T+0
scope SLIDE 02 : WHAT AGENT OPERATIONS ACTUALLY COVERS What agent operations covers: SLOs, on-call, incident response, cost management
Reliability isn't luck : it's engineering

Day-2 discipline for a fleet, not a demo

Deploying an agent version into production is just the beginning. Operations ensures its ongoing health through defined severities, on-call rotations, incident response with runbooks, postmortems for learning, and proactive cost and capacity planning.

AgentOps was created to help organizations make intentional trade-offs between reliability and speed when implementing agent changes, as without service-level objectives, each change poses a risk with unpredictable outcomes.

◷ T+0
SLOs SLIDE 03 : SLOs FOR SYSTEMS THAT VARY BY DESIGN Composite SLOs and statistical anomaly detection for agent reliability
A hard threshold assumes the wrong thing

Learn what normal looks like, then watch for drift

Conventional SRE thresholds are based on the assumption that a metric remains consistent under normal conditions. However, agent outputs are intentionally diverse, even when functioning properly, making a fixed threshold either too sensitive or not sensitive enough. An alternative approach is statistical anomaly detection, where the system learns the typical behavior of an agent and only alerts when it deviates beyond preset boundaries, rather than reacting to individual abnormal outputs.

Composite SLOs bring together various signals to establish a unified reliability goal, as a single metric cannot accurately gauge the overall health of an agent fleet. Reliability, quality, speed, and cost each require their own target, integrated rather than monitored separately on different dashboards.

◷ T+0
When something breaks

The five-phase incident response loop

One faulty prompt can spread through numerous runs before anyone notices - the loop is in place to prevent it from escalating.

incident SLIDE 04 : DETECTION → CONTAINMENT → ERADICATION → RECOVERY → POSTMORTEM Five-phase incident response loop for agentic workflows
Every eradication action should be reversible

Roll it back, but keep the evidence

Revert a faulty prompt, but save the faulty version for analysis. Secure a model, but document the failure on the reverted version for a safe retry. Deleting evidence eliminates the opportunity to learn from incidents.

01
Detection

Statistical signals flag drift before it's visible to a user.

02
Containment

Stop the bleeding : quarantine the failing component.

03
Eradication

Remove the cause reversibly, preserving evidence for later.

04
Recovery

Unpause in measured tranches, verifying between each one.

05
Postmortem

Discover the underlying reason, not just the apparent symptom.

◷ T+0
detection SLIDE 05 : THE SIGNALS THAT CATCH AGENT FAILURES Detection signals that catch agent-specific failures classical monitoring misses
Each minute of detection incurs a significant cost.

Watching for agent failures, not web failures

Traditional web monitoring tools are not equipped to detect agent failures, leading to potential cost increases of up to fifty times in just one hour, even when traditional error rates remain unchanged. A more comprehensive detection system is needed to monitor agent decision deviations, escalate issues to human operators, assess tool selection consistency in workflows, and identify outliers at the tenant level.

The existence of these signals relies on trace coverage at tool-call granularity and cost attribution per workflow, with the detection of the instrumentation layer depending entirely on this.

◷ T+0
How much should the agent do alone

Autonomy is a property of the action, not the agent

One agent may function with varying levels of autonomy depending on the task at hand – cache invalidation, service restarts, and database changes all warrant different levels of control.

on-call agents SLIDE 06 : THREE-TIER AUTONOMY MODEL Three-tier autonomy model for AI agents in on-call incident response
On-call isn't going away : it's changing shape

Not replacing on-call, augmenting it

Being on-call involves more than following a runbook; it requires identifying gaps in the runbook and making quick decisions under stress. Agents are not equipped to handle this aspect effectively, so a practical autonomy approach limits independent actions to a specific, pre-approved list of common, low-risk scenarios.

Advisory

Suggests, human decides

The agent scans logs to identify incidents that match a runbook signature and recommends a remediation, which is then approved by a human operator.

Approval-gated

Proposes, human confirms

Agent suggests a precise action - 'undo the recent deployment?' - and displays the exact modifications. It waits for a direct approval from a human before taking any action.

Conditional autonomy

Acts, within a narrow scope

The agent operates independently, expanding its scope cautiously as it proves the safety of restarting a crash-looped stateless service with a small set of pre-approved, low-risk patterns.

◷ T+0
runbooks SLIDE 07 : FROM WIKI PAGES TO MACHINE-READABLE PLAYS Converting static runbooks into structured, agent-readable playbooks
A dusty wiki is too slow for an incident

Structured plays an agent can actually execute

A static runbook page may be sufficient for a person reading at their leisure, but incidents cannot wait. To begin the migration process, start small by analyzing the last few months of incidents, identifying recurring patterns, and transforming the top three to five most frequent, low-risk incidents into organized steps that an agent can practice before implementation.

Integrate those plays into the incident tool initially as a recommendation for human execution, gradually progressing to one-click or automatic execution for the safest actions once the data confirms they align with the team's intended approach.

◷ T+0
cost & capacity SLIDE 08 : QUOTAS, BUDGETS, AND COMPOUNDING FORECASTS Cost and capacity operations: quotas, budgets, and forecasting for agent fleets
Cost is an operational metric, not a finance surprise

Budget per agent, alert on the surge before the invoice

Establish quotas and cost budgets for each agent and user, consistently track token and dollar costs per task, and connect the generated dashboards directly to SLOs instead of assessing them monthly. Immediate alerting on any cost surges prevents a minor issue from becoming a major problem that goes unnoticed for weeks.

Capacity planning is further complicated by the incorporation of agentic forecasting models that take into account operational telemetry such as incident frequency, vendor lead times, and request seasonality, which are often overlooked by traditional forecasting methods, leading to continuous improvement in quarterly predictions.

◷ T+0
postmortem SLIDE 09 : POSTMORTEMS THAT FIND THE REAL CAUSE Postmortem discipline that avoids agent-blame and finds the real root cause
"The agent hallucinated" is a symptom, not a cause

Force the postmortem to look at the system

In an agent postmortem, the most frequent failure mode is attributing the root cause to 'the agent hallucinated' and then closing the ticket without identifying the true reason. This is typically due to a lack of guardrails, incomplete evaluation coverage, errors in context engineering, or discrepancies between a tool's schema and the agent's expectations.

A postmortem template with clear prompts at every step ensures a comprehensive analysis of the system, rather than relying on the writer to uncover hidden issues independently.

outcomes SLIDE 10 : THE BORING 80%, HANDLED Summary: agents handling the boring 80% of operations, humans keeping the interesting 20%
The result worth measuring

MTTR halved, without a different runbook

Teams that effectively implement incident-response augmentation can expect a significant decrease in mean-time-to-resolution within a short period of deployment, without any alterations to the severity matrix or escalation policy. The improvement is mainly attributed to eliminating slow manual tasks within the existing runbook, such as reading logs, locating the necessary information, and drafting updates, rather than implementing a completely new process.

Instead of reducing the number of people involved, new positions are created: an Incident Commander to oversee the interaction between humans and technology in real-time during emergencies, and a Capacity Analyst whose predictions significantly impact operational decisions. The objective is not to decrease human involvement in operations, but to ensure that resources are utilized effectively.

Frequently asked

Agent operations FAQ

Deploying a new agent version safely into production involves staged rollout gates, while ensuring ongoing system health through various operational tasks such as SLOs, on-call support, incident response, runbooks, postmortems, and cost and capacity management. This is essential for the sustained success of the system beyond just the initial release.

As agent outputs are intentionally diverse even under normal operation, using a fixed threshold can be either too chaotic or too limited in detecting actual issues. Statistical anomaly detection, on the other hand, focuses on learning an agent's typical behavior and notifies of significant deviations from that baseline, rather than comparing each output to a single predetermined value.

Detection involves identifying statistical signals indicating drift, containment focuses on halting the spread of failure, eradication entails removing the reversible cause while preserving evidence, recovery entails resuming operations in verified tranches, and postmortem involves identifying the true root cause rather than just the visible symptoms.

Autonomy should be limited to a specific, low-risk set of frequent actions, with each action having its own scope rather than being applied universally across all agents. For example, an agent may be able to independently perform cache invalidation but would need human confirmation for service restarts and manual review for any database-related tasks.

The reason for a hallucination is often a specific and fixable issue, such as a missing guardrail or a gap in evaluation coverage, rather than just the symptom of the failure.

Implementing agentic triage, log synthesis, and root-cause hypothesis generation typically results in a 50% reduction in mean-time-to-resolution within a quarter for teams, without altering the severity matrix or escalation policy. This enhancement is achieved by streamlining the cognitive processes within the current runbook, rather than introducing a new procedure.

Get started

Ready to build your agent operations playbook?

Develop composite SLOs prior to your next incident, transform your top three most frequently used runbooks into organized plays, and establish autonomy for each action from the start.