Embracing Observability for the AI era at re:Invent 2026

As organizations continue to scale their cloud presence and bring AI agents into production, effective operations become increasingly critical for success. The AWS re:Invent 2026 Cloud Operations track brings together industry experts, AWS leaders, and customers to share how they modernize monitoring and observability for applications and the AI agents now running alongside them.

This post walks you through the key themes of the Monitoring and Observability lineup and highlights the sessions that will help you transform your cloud operations strategy, with AI-assisted investigation in Amazon CloudWatch as the connective thread.

Plan your Monitoring and Observability track experience

With 25 sessions spanning six key themes, the Monitoring and Observability lineup offers something for everyone, from hands-on workshops to expert-level discussions. To make the most of your re:Invent experience, we recommend:

1. Focus on your priorities: select sessions that align with your organization’s immediate operational challenges.
2. Mix formats: combine lecture-style sessions with interactive workshops and builders’ sessions.
3. Plan for skill development: choose sessions that match your current skill level and those that stretch your capabilities.
4. Reserve early: popular sessions fill up quickly, so reserve your spot as soon as registration opens.

Observe AI agents in production

The future of cloud operations is AI-driven, and these sessions show how to observe agents that reason, call tools, and act on their own:

COP304 | Observability and operations in the AI era | Breakout Session

Location: Tuesday Dec 1 4:00 PM–5:00 PM PST | Wynn/Encore · Level 1 · Encore Ballroom 5

Generative AI broke the assumptions traditional monitoring was built on, and teams struggle to keep non-deterministic systems predictable, explainable, and affordable. This session opens on that challenge, then shows how the newest observability capabilities give you visibility into model quality, cost, and agent behavior across your AI stack, using traces, spans, and confidence scoring, illustrated with real customer patterns.

Who should attend: For SREs and platform teams operating generative AI workloads who need visibility into model quality, cost, and agent behavior.

COP309 | I didn’t know my AI agent was doing that! | Builders’ Session

Location: Monday Nov 30 1:30 PM PST | Wynn/Encore · Convention Promenade · Latour 7

AI agents are non-deterministic: they can burn tokens, stall, or drift into wrong behavior with no warning, and teams often have no visibility until something breaks. In this hands-on builders session, you will use Amazon CloudWatch to observe live agents from both the console and your IDE: trace what an agent actually did, spot failures, and surface anomalous behavior. Working with real agents, a production agent and a custom chatbot, you build the detection and alerting yourself and leave able to apply it to the agents you run in production.

Who should attend: For engineers running agents in production who want hands-on tracing and anomaly detection from the console and IDE.

COP318 | Catch it before production: Instrument your AI agents | Workshop

Location: Monday Nov 30 12:00 PM–2:00 PM PST | Wynn/Encore · Level 1 · Brahms 4

You need visibility into your AI workloads before it reaches production, not after it fails, yet most teams have no telemetry on their agents at all. In this workshop, you instrument a live AI agent end to end, capturing token cost, latency, tool calls, and reasoning traces. You make the instrumentation decisions, build the dashboards and anomaly alerts. Leave with a working setup you can apply to your own agentic workloads.

Who should attend: For teams putting their first AI agents into production who need to instrument them before launch.

COP319 | Observe AI agents beyond your perimeter | Builders’ Session

Location: Tuesday Dec 1 1:00 PM PST | Caesars Forum · Level 1 · Forum 120 · Content Hub · Builders’ Session 2

Your developers are burning tokens on AI coding agents with zero visibility. Claude Code and OpenAI Codex run on laptops, outside your perimeter, with no telemetry reaching your observability backend. In this session, we fix that. Using native OpenTelemetry export and Amazon CloudWatch’s OTLP endpoint with a single bearer token, with no collectors, no sidecars, no credential wiring, we ship metrics (token usage, cost, productivity), logs (tool decisions, errors), and traces (agent reasoning chains) directly from developer machines. Then we build a live dashboard with per-team cost attribution and anomaly alerting, all queryable with PromQL. Walk out ready to deploy in minutes.

Who should attend: For platform and security teams who need telemetry on AI coding agents running outside their perimeter.

COP325 | Observability patterns for agentic AI in production | Chalk Talk

Location: Monday Nov 30 1:00 PM–2:00 PM PST | Wynn/Encore · Convention Promenade · Latour 5

AI agents are non-deterministic, multi-step, and make autonomous decisions and traditional APM isn’t enough. Explore observability patterns for agentic workloads: reasoning chain tracing, token cost attribution, guardrail monitoring, and multi-agent coordination. We’ll whiteboard architectures in real time. Bring your toughest production agent monitoring questions and leave with actionable patterns.

Who should attend: For architects and SREs designing observability for agentic workloads who want proven patterns.

COP337 | Wrong, Not Broken, & Why your Observability should know the difference | Breakout Session

Location: Wednesday Dec 2 12:00 PM–1:00 PM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Purple Theater

Agentic applications need different access patterns and tooling than traditional apps. Teams often stitch together several tools and act as the human orchestrator just to answer a basic question: is my agent doing what the business needs? That gets harder when the agent runs on one cloud and calls models or APIs on another. In this session, we show how CloudWatch Omni brings AI agents and applications across clouds into one view, accelerating root cause whether it is due to performance, availability or response quality.

Who should attend: For teams operating agents and apps across multiple clouds who need one view for root-cause analysis.

COP349 | Build a per-tenant observability stack for Agentic SaaS | Builders’ Session

Location: Monday Nov 30 4:30 PM PST | Wynn/Encore · Convention Promenade · Latour 7

Multi-tenant ISVs need per-tenant observability that traditional dashboards can’t deliver — aggregated metrics hide the tenant causing the problem, and shared traces don’t tell you whose agent misbehaved. In this session, build a per-tenant observability stack for agents. You’ll instrument agent invocations with tenant-aware traces and session replay, build tenant-scoped dashboards that surface noisy neighbors and latency outliers, and configure alerting that fires on per-tenant thresholds rather than aggregate signals. By the end, you’ll have an observability setup that scales with your tenant count without a separate data pipeline — and patterns to extend it as your agent architecture grows.

Who should attend: For ISVs and multi-tenant SaaS builders who need per-tenant visibility into agent behavior.

COP401 | Building eval-driven observability for AI agents | Code Talk

Location: Monday Nov 30 4:30 PM–5:30 PM PST | Caesars Forum · Level 1 · Forum 120 · Content Hub · Code Talk

Multi-agent systems fail in ways your dashboards cannot show: one agent hands off bad context, another retries, and the API still returns 200. In this live coding session, start with a multi-agent app on LangGraph, Strands, and CrewAI, and instrument it with OpenTelemetry and OpenInference. Trace it from the first prototype run, evaluate agents on demand with LLM-as-judge evaluators, A/B test prompt and model variants before promoting to production, pull production traces into a versioned dataset, reproduce the failure locally, and ship the fix. All from your IDE, without switching tools.

Who should attend: For developers building multi-agent systems who want eval-driven observability in their workflow.

COP407 | Build-time context for faster agentic incident response | Code Talk

Location: Thursday Dec 3 10:30 AM–11:30 AM PST | Wynn/Encore · Convention Promenade · Margaux 1

AI agents can fix production issues, but only with the right context. In this talk, an AI agent builds a feature and, in the same pull request, creates the context it needs on call: instrumentation, alarms, dashboards, and skills, all as code. Then you break the live app. Watch an agent use those skills with Amazon CloudWatch Omni and AWS DevOps Agent to triage, find the root cause, and ship the fix. Leave with patterns that give agents the context to operate what they build.

Who should attend: For teams that want AI agents to operate what they build, with context created at build time.

INV521 | Teaching Telemetry to Think: From Raw Signals to Root Cause | Chalk Talk

Location: Wednesday Dec 2 1:00 PM–2:00 PM PST | Wynn/Encore · Convention Promenade · Latour 5

An AI agent investigating an incident starts cold: no sense of normal, no topology, no causal history. It rebuilds context from raw documents every time. What if the telemetry understood itself first? We’ll examine an intelligence layer for agentic observability, where streaming detectors and a live service graph distill billions of signals into incident observations, then fine-tuned reasoning models, built on open-weight LLMs, correlate them into ranked root causes. Whiteboard with us on the hard parts: what to teach a model versus retrieve at inference, how domain fine-tuning beats frontier scale, and where reasoning models still fail.

Who should attend: For architects and researchers interested in how reasoning models turn telemetry into ranked root causes.

Build an observability strategy for the AI era

Set the direction for observability as AI reshapes how operations teams work:

COP335 | Evolve your observability for the agentic era | Breakout Session

Location: Tuesday Dec 1 11:30 AM–12:30 PM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Blue Theater

Your agent passed every test last week. Today the same prompt returns a different, pricier answer, and nobody can say why. Agents drift as models and prompts change beneath them, and a green dashboard won’t catch it. This session covers what evaluations (“evals”) are: automated checks that score an agent’s output for quality, accuracy, and cost. You’ll learn how an eval-driven loop catches that drift on live traffic, traces it to the change that caused it, compares models and prompts, and attributes spend down to the agent and user, so regressions get blocked before they reach production. Then see how Amazon CloudWatch Omni does this across AWS and other clouds, covering your agents and applications in one place.

Who should attend: For operations teams consolidating logs, metrics, and traces for troubleshooting at scale.

COP340 | Best practices for building an observability strategy in the age of AI | Chalk Talk

Location: Thursday Dec 3 8:30 AM–9:30 AM PST | Caesars Forum · Level 1 · Academy 409

Most observability strategies were built for a pre-AI world and are no longer enough. Your data foundation remains critical, but the way you use that data is changing fast. In this chalk talk, we explore what stays the same, what must evolve, and what’s new as AI reshapes observability. We’ll cover the shift from static dashboards to dynamic experiences, from reactive alarms to AI-driven insights, and the new questions you need to answer, like how well your agents are performing. Join us as we share what AWS is learning as we embrace AI-led observability.

Who should attend: For leaders and architects rethinking observability strategy for the AI era.

Standardize on OpenTelemetry

Open standards are the foundation of observability on AWS, and these sessions put them to work:

COP306 | Cut through Kubernetes signal overload with unified observability | Builders’ Session

Location: Wednesday Dec 2 10:00 AM PST | Caesars Forum · Level 1 · Forum 120 · Content Hub · Builders’ Session 2

Kubernetes at scale buries teams in high-cardinality signals that traditional pipelines truncate or split across separate backends, so the answer you need is never in one place. In this hands-on builders session, you send OpenTelemetry metrics and logs from an EKS cluster into Amazon CloudWatch, deploy Container Insights, and decide which signals actually matter. You use automatic resource enrichment to build your own dashboards and alerts, deciding what to alert on and why rather than running pre-made scripts. You leave knowing how to cut through high-cardinality noise and set up alerts that drive action.

Who should attend: For Kubernetes and platform teams drowning in high-cardinality EKS signals.

COP308 | Turn raw telemetry into answers with OpenTelemetry | Builders’ Session

Location: Wednesday Dec 2 11:30 AM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Builders’ Session 1

Raw telemetry arrives noisy, unlabeled, and hard to query, so teams waste time cleaning data before they can act. In this hands-on builders session, you set up OpenTelemetry for metrics, logs, and traces from a live sample application, then use Amazon CloudWatch pipeline processors to enrich and normalize the data in flight. You build the enrichment rules, explore results with natural language query, and leave with a reusable pattern for turning raw signals into analysis-ready telemetry.

Who should attend: For engineers standardizing instrumentation who want to enrich telemetry in flight.

COP336 | Elevate your observability with OpenTelemetry on AWS feat. Genesys | Breakout Session

Location: Wednesday Dec 2 2:30 PM–3:30 PM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · White Theater

Getting full visibility across your stack shouldn’t mean stitching together disconnected tools and proprietary formats. When you standardize on OpenTelemetry with AWS, you get rich, unified telemetry from metrics, logs, and traces across your containers, Kubernetes clusters, AI workloads, and everything in between. Collect deep telemetry using OTLP ingestion, PromQL querying, and automatic enrichment with Amazon CloudWatch’s native OpenTelemetry. No fragmented data. No vendor lock-in. Just one telemetry foundation that gives you the context you need to understand what’s happening across your entire environment. Learn how to achieve richer data, unified observability, and a production-ready architecture that scales with you.

Who should attend: For teams standardizing on OpenTelemetry across containers, Kubernetes, and AI workloads.

COP343 | Instrument a service with OpenTelemetry | Lightning Talk

Location: Wednesday Dec 2 2:30 PM–2:50 PM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Lightning Theater

In this live demo, we take an uninstrumented service from zero observability to correlated metrics, logs, and traces in minutes using Open Telemetry and Amazon CloudWatch. Walk through the shortest path to production-grade instrumentation: SDK configuration, collector setup, signal routing, and surfacing the first actionable insight.

Who should attend: For engineers new to OpenTelemetry who want the shortest path to instrument a service.

Turn logs into intelligence

Move from storing logs to acting on them across every account and region:

COP310 | Automate threat detection and response across your logs | Builders’ Session

Location: Monday Nov 30 1:30 PM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Builders’ Session 2

Security teams cannot manually hunt through every log for the one event that matters, and threats move faster than analysts can query. In this hands-on builders session, you will build automated detection and response on your security telemetry: metric filters and anomaly detectors that flag suspicious activity, findings routed to alarms that page the right people, and automated remediation that acts before an analyst logs in. After tuning detections to cut false positives, walk away with a pipeline that watches your logs so your team does not have to.

Who should attend: For security and operations teams automating detection and response on log data.

COP339 | Follow that log from pipeline to intelligence | Chalk Talk

Location: Thursday Dec 3 1:00 PM–2:00 PM PST | Wynn/Encore · Convention Promenade · Montrachet 1

You’ve got logs flowing from everywhere, but when an issue spans multiple accounts or regions, you’re stuck jumping between consoles trying to piece together what happened. In this chalk talk, we’ll whiteboard the end-to-end path a log takes from ingestion pipeline to a unified data layer to log analytics. We’ll break down where visibility gaps form and why. Learn how you can connect these layers to trace a problem across your entire environment without losing context.

Who should attend: For teams tracing issues across accounts and regions who want to close log visibility gaps.

Trace performance across the full stack

When an application slows down, these sessions help you find the cause across every layer:

COP322 | Where’s the slowdown? Build observability from frontend to databases | Workshop

Location: Monday Nov 30 12:00 PM–2:00 PM PST | MGM Grand · Level 1 · Grand 120

When an application slows down, the cause can hide anywhere from the frontend to the database, and siloed tools make teams guess. In this workshop, you will build full-stack observability across application, infrastructure, and database layers, bringing RDS and Aurora signals and database logs into one store and correlating them with audit activity. You decide what to trace and build the dashboards that connect a symptom to its root cause. Leave with skills to trace any slowdown across layers in a single pane.

Who should attend: For full-stack and database teams diagnosing slowdowns across application and data layers.

COP338 | High latency! Is it the code, the container, or the host? | Chalk Talk

Location: Monday Nov 30 10:00 AM–11:00 AM PST | MGM Grand · Level 3 · Premier 311

Your app is slow, but is it the code, the container, or the underlying host? When your stack spans clouds, hybrid, and on-prem, correlating application performance to infrastructure health means jumping between tools and guessing at root cause. Join this session as we walk through how to map application signals such as traces, errors, and latency directly to the infrastructure, regardless of where it runs. You’ll see how to pinpoint whether the problem is your app or the compute it sits on, without stitching dashboards together manually.

Who should attend: For SREs correlating application performance to infrastructure across cloud, hybrid, and on-prem.

COP342 | Optimize container observability to reduce cost | Chalk Talk

Location: Monday Nov 30 4:30 PM–5:30 PM PST | MGM Grand · Level 1 · Boulevard 165

At scale, container environments generate metrics and logs that increase both cost and noise for engineering teams. This session explores architectural strategies for managing observability data across containerized workloads using OpenTelemetry. Learn how to control metric cardinality, design sampling policies that preserve critical signals, and evaluate managed backend tradeoffs. You’ll leave with a framework for building cost-aware container observability.

Who should attend: For container platform teams who need to control observability cost at scale.

COP408 | Correlate database query latency to user impact | Code Talk

Location: Friday Dec 4 9:00 AM–10:00 AM PST | Wynn/Encore · Convention Promenade · Margaux 1

Diagnosing performance degradation across distributed applications requires more than isolated database or application metrics. It demands correlated, end-to-end telemetry. This code talk demonstrates a scenario that unifies database performance signals with application-layer telemetry using OpenTelemetry and Amazon CloudWatch. Learn how to implement trace-to-query correlation, design log aggregation patterns for accelerated root cause analysis, and build metric relationships that surface bottlenecks across the full stack. You’ll be able to connect a slow database query to its downstream user impact.

Who should attend: For database and application teams connecting query latency to user impact.

COP409 | Pinpoint latency in microservices with distributed tracing | Code Talk

Location: Monday Nov 30 1:30 PM–2:30 PM PST | Wynn/Encore · Convention Promenade · Margaux 1

Diagnosing latency across a high-scale Amazon EKS environment is complex when pods scale dynamically and trace context fails to propagate across service boundaries. In this code talk, we instrument containerized microservices with OpenTelemetry, propagate context across service hops, and correlate spans, metrics, and logs in Amazon CloudWatch. Live coding shows how a single slow request reveals the exact container and code path responsible for degradation, using targeted span attributes and exemplars to make Kubernetes-scale traces navigable. You’ll also see the trade-offs behind sampling strategies and collector deployment topologies that affect overhead, fidelity, and cost at production scale.

Who should attend: For microservices teams pinpointing latency with distributed tracing on EKS.

Run open source observability at managed scale

Keep the open source tools your teams rely on while AWS runs the backends:

COP307 | Keep your open source observability, move to managed scale | Builders’ Session

Location: Wednesday Dec 2 9:00 AM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Builders’ Session 2

Teams standardized on Prometheus, Grafana, and OpenSearch should not have to abandon their tools to get scale and lower operational overhead. In this hands-on builders session, you will learn about instrumentation and collection powered by OpenTelemetry, and explore that telemetry into managed open source services: ship metrics to PromQL in Amazon Managed Service for Prometheus, logs and traces in Amazon OpenSearch Service, and correlate all signals in Amazon Managed Grafana. Own your telemetry pipelines end-to-end with open standards and AWS managed services.

Who should attend: For teams on Prometheus, Grafana, and OpenSearch who want managed scale without changing tools.

COP410 | Build a fully managed open source observability stack on AWS | Code Talk

Location: Friday Dec 4 8:30 AM–9:30 AM PST | Wynn/Encore · Upper Convention Promenade · Cristal 2 · Content Hub · Code Talk

You want open standards without the operational burden of self-managing backends. In this code talk, we live-code a complete open source observability architecture: instrument a service with OpenTelemetry (auto and manual), configure the OTel Collector pipeline, and route metrics to Amazon Managed Service for Prometheus, logs and traces to Amazon OpenSearch Service. Walk through each integration decision from SDK selection to collector topology, and leave with a working, production-ready stack built entirely on open source standards with zero infrastructure to manage.

Who should attend: For teams building a fully managed open source observability stack on AWS.

Conclusion

The Monitoring and Observability lineup at re:Invent 2026 spans observing AI agents, standardizing on OpenTelemetry, turning logs into intelligence, tracing performance across the full stack, and running open source at managed scale. Review the themes, reserve your seats early, and tell us in the comments which observability challenges you most want to solve. We will see you in Las Vegas.

Haven’t registered? There is still time to attend. Register through the re:Invent portal.