AWS Builder Center
Learn from the Builders Who've Done It: The re:Invent 2026 Community Track

Learn from the Builders Who've Done It: The re:Invent 2026 Community Track

Your guide to the re:Invent 2026 Community Track, led by AWS Heroes, Community Builders, and User Group Leaders.

Learn From the Builders Who've Done It: The re:Invent 2026 Community Track

The Community (COM) track at re:Invent is where you hear from the people who build on AWS every day. These sessions are led by some of my favorite people, AWS Heroes, AWS Community Builders, and AWS User Group Leaders: practitioners who run production workloads, maintain open source projects, organize local meetups, and share what they've learned through blogs, videos, talks, and so much more.
This year the track features 43 sessions, and 10 of them run twice, so you have more chances to fit them into your schedule. None of these are product pitches. Each talk is a practitioner's account of a real system: what they built, what broke, the trade-offs they made, and the patterns you can take back to your team.
Below is the full lineup, grouped by theme. To add sessions to your agenda, search "COM" in the  re:Invent catalog . Each speaker's Builder ID links to their AWS Builder Center profile, so you can follow them and keep learning after re:Invent.

Serverless & Compute

COM303 | Lambda for every scale and workload shape (Breakout, 300)
Liz Fong-Jones ( @lizthegrey ) · Wed, Dec 2, 1:00 PM · Caesars Forum
Honeycomb spends millions of dollars a year on Lambda. We outgrew Lambda Functions, not Lambda: Managed Instances now absorbs steady-state baseline traffic at roughly 7.5x lower cost than Functions, MicroVMs takes spikes of load tolerating 1.5 seconds of added latency, and Functions still catches bursty spillover. We'll walk through the observability and benchmarks that drove each cutover, and the routing logic deciding which tier a query lands on before it executes. You'll learn how to scale your serverless workloads multiple orders of magnitude without linear cost.
COM340 | Lambda MicroVMs: The Missing Serverless Primitive for Multi-Tenant Agents (Lightning, 300)
Darshit Pandya ( @dp99 ) · Wed, Dec 2, 12:30 PM · Wynn/Encore
Lambda Functions are stateless and time out after 15 minutes. Fargate tasks are expensive to keep warm. When building multi-tenant agent platforms, I needed per-tenant isolation without per-tenant cost. Lambda MicroVMs solve this: Firecracker VM-level isolation, 8-hour stateful execution, auto-suspend when idle, instant resume on next request. This session walks through the architecture I built, replacing always-on containers with MicroVMs as per-tenant agent sandboxes. I'll share the tradeoffs, what surprised me, and a deployable pattern you can take home and ship.
COM344 | Now serverless has servers. What now? (Lightning, 300)
Evandro Pires ( @epiresdasilva ) · Mon, Nov 30, 6:00 PM · Venetian, Expo Builder Theater
I spent years advocating for serverless, but as a CTO, I could not always see a viable path to bring it into certain business scenarios. In this talk, I share what has changed with the new AWS capabilities, why this can change the game for larger companies, and, along the way, I explore the concepts and how these services work in more depth to help us think about architecture with more clarity.
COM202 | Rebuilding a Platform's Foundation on DynamoDB Streams and Lambda (Lightning, 200)
Awedis Keofteian ( @awedis ) · Mon, Nov 30, 3:30 PM · Wynn/Encore
Our platform had single points of failure baked into its original design. In this session, I walk through how we closed those gaps across all five AWS Well-Architected pillars: moving event-driven workloads onto DynamoDB Streams and Lambda, cutting costs through smarter data tiering, and building the observability needed to trust the new design under real operational pressure.
COM314 | Beyond the GPU count: Engineering inference at scale (Breakout, 300)
Mai Nishitani ( @electrokat ) · Thu, Dec 3, 11:30 AM · Caesars Forum
In a world where everyone is fighting for GPUs, picking the wrong server type is one of the most expensive mistakes an engineering team makes at scale. This session shares lessons running GPU inference across AWS and bare metal connected privately. You'll see how EKS Hybrid Nodes provides a single scheduling plane, failure modes, and how to use AI to accelerate decisions without replacing engineering judgement. Leave with a cloud and bare metal architectural pattern and a GPU selection framework covering workload fit and server tradeoffs.
COM345 | Designing for Operability: Catching Performance Gaps Before Scale (Lightning, 300)
Jofia Jose Prakash ( @jofia ) · Tue, Dec 1, 4:30 PM · Wynn/Encore
A training workload had plenty of GPU capacity, yet utilization remained low and training time kept increasing. The GPUs were healthy. The real bottleneck was upstream: the data pipeline could not feed them fast enough, leaving expensive accelerators waiting for work.
This talk walks through how we traced that utilization gap, found where throughput was being lost, and identified the signals that could have exposed the problem earlier. Through this production experience, we’ll look at what it means to design for operability before scale, when performance, utilization, and cost problems are still easier to fix.

Containers & Kubernetes

COM404 | Scaling Multi-Cluster EKS Without Burning Your Infrastructure Budget (Chalk talk, 400)
Sagar Utekar ( @sagar2366 )
  • Wed, Dec 2, 1:30 PM · MGM Grand, Boulevard 155
  • Thu, Dec 3, 12:00 PM · MGM Grand, Premier 309
Managing hundreds of Amazon EKS clusters breaks standard multi-tenant architectures under the weight of shared control planes, noisy neighbors, and runaway cross-AZ data transfer costs. We will dissect the architectural trade-offs between monolithic clusters and localized micro-clusters, mapping out exact blast-radius constraints. Bring your worst infrastructure bottlenecks; we will openly debate cluster life-cycle automation, shared VPC patterns, and cost-allocation frameworks, leaving you with a battle-tested blueprint to optimize fleet operations on Monday morning.
COM323 | From Fragmented Infra to Golden Path: An Internal Developer Platform on EKS (Code talk, 300)
Johannes Koch ( @lockhead ) · Wed, Dec 2, 1:30 PM · Caesars Forum
Every team wanted to ship faster, but each was reinventing the same wheels: clusters, pipelines, infrastructure, access. So we built an internal developer platform on Amazon EKS — one golden path that lets product teams go from idea to running service without becoming Kubernetes experts. This is the honest story of that journey. We'll show how we used EKS as the foundation, Crossplane to provision cloud infrastructure through Kubernetes APIs, and Argo CD to drive everything through GitOps. You'll leave with a realistic map of the terrain: what worked, what hurt, and what we'd do differently.
COM329 | Sandboxing AI Agents on EKS: Isolation That Scales (Code talk, 300)
Adit Modi ( @adit ), Ishan Modi ( @ishanmodi ) · Tue, Dec 1, 4:30 PM · Caesars Forum
AI agents execute untrusted code — one prompt injection away from data exfiltration. This session covers building graduated isolation for autonomous agents on EKS using the agent-sandbox controller. You'll learn why runtime tiers (runc → gVisor → Firecracker) map to different threat models, how SandboxClaim resources manage ephemeral agent lifecycle, and Cilium FQDN egress filtering that prevents exfiltration without blocking legitimate tool use. Leave with a production-ready isolation framework.
COM320 | Lights, Camera, Fargate: A Reality-TV Startup Hardened for Scale on AWS (Breakout, 300)
Brian H. Hough ( @brianhhough ), Erik Hanchett · Fri, Dec 4, 8:30 AM · Venetian
In six weeks on reality TV coding show The Full Stack, we built DataSwarm — a data-observability solution shipped as an open-source CLI with cloud-hosted data analysis and remediation. Great demo, wrong architecture. When real users arrived, the demo cracked. So we pivoted. Long-running agents moved off Lambda onto ephemeral ECS Fargate tasks. API keys swapped for firewalled Amazon Bedrock, feature-branched CI/CD with dev-to-prod promotion, and Next.js behind CloudFront, S3, and Lambda@Edge. Leave knowing how to tell demo-grade architectures from customer-grade ones — and when to make the jump.

Distributed Architecture & Resilience

COM304 | Hidden trade-offs in distributed architectures (Breakout, 300)
Anahit Pogosova ( @anahit ) · Wed, Dec 2, 4:00 PM · MGM Grand, Grand 117
Asynchronous event-driven architectures enable loosely coupled, scalable, and resilient systems. But the devil is in the details, and hidden complexity emerges when components interact under real-world pressure. Consistency, latency, failure handling, and cost aren't determined by architecture diagrams but by the trade-offs behind the integration patterns and services you choose.
Through a real-world case study using AWS Lambda, Amazon DynamoDB, DynamoDB Streams, & AWS Step Functions, you'll investigate unexpected system behavior and uncover the trade-offs behind it. You'll leave with a sharper intuition to see beyond cloud abstractions and make more intentional architectural decisions for resilient systems.
COM309 | A DevOps Approach to Recover at Scale from Disaster or Ransomware (Breakout, 300)
Samantha Eliot ( @setheliot ) · Thu, Dec 3, 8:30 AM · Caesars Forum
Recovering your cloud workload completely after a regional event or ransomware attack is not a checklist; it demands a culture of resilience that turns individual craft into reliable recovery. That culture is built through DevOps. This session covers the technical requirements for recovering cloud workloads at scale, and the DevOps practices that make those requirements actually work under pressure. You'll leave with practical knowledge to drive full recovery in the face of the worst disasters.

Security, Identity & Compliance

COM402 | Practical PKI for On-Prem Devices with IAM Roles Anywhere (Breakout, 400)
Ben Bridts ( @benbridts ) · Wed, Dec 2, 10:00 AM · Wynn/Encore, Encore Ballroom 5
Secure authentication of your humans and machines is a challenge every organization faces, especially if your proof of identity is not already in AWS. In this session, we'll talk you through multiple ways to set up or use PKI - such as Belgium's eID or a hardware key - to securely access AWS with phishing resistant authentication built using IAM Roles Anywhere. This will include live demos & code samples. By the end of the session we will stop worrying about plain text credential storage because we can use trusted hardware instead.
COM315 | Beyond the Docs: Operating Scoped VPC Endpoints in Real Environments (Chalk talk, 300)
Meg Ashby ( @megashby1 ), Kenneth Attard ( @kattard ) · Thu, Dec 3, 9:00 AM · Caesars Forum, Forum 114
Scoped VPC endpoint policies look simple, until production traffic proves otherwise. In this interactive chalk talk, we'll walk through real-world failures we've encountered while deploying account- and org-scoped endpoint policies and trace the identity flows, auth chains, and hidden dependencies behind them. Together, we'll explore a key lesson: endpoint policies evaluate the request that's actually made, not the request you think you're making.
COM302 | The judgment gap: when Terraform scans clean but fails the audit (Chalk talk, 300)
Anton Babenko ( @antonbabenko )
  • Mon, Nov 30, 1:30 PM · Caesars Forum, Academy 414
  • Thu, Dec 3, 2:30 PM · Caesars Forum, Forum 103
Checkov, trivy, and Prowler check rules someone already wrote. SOC 2 CC6.1 is a sentence open to interpretation. Between them sits the judgment gap, where AI-generated Terraform for S3, CloudFront, API Gateway, and Lambda often lands. We map the gap on the whiteboard together: take real controls from SOC 2 and PCI DSS, set them against what scanner rules verify, and mark what is left over. Then we discuss what closes the remainder: better authoring-side skills, human review, multi-model review, or accepting the risk. Bring a control your team owns; we trace where its enforcement lives, or does not.
COM403 | One-Way Doors: Launching a Regulated Financial Exchange on AWS in 2026 (Chalk talk, 400)
Margaret Valtierra ( @margaretvaltierra ), Jane Ginsburg · Mon, Nov 30, 4:30 PM · Wynn/Encore, La Tache 2
AWS turns 20 this year. Starting fresh in 2026 means navigating hundreds of services with no clear map for which decisions are one-way doors. I'll share how we built a CFTC-regulated exchange & clearinghouse from the AWS Organizations up. The full security stack (Access Analyzer, Config, GuardDuty, Macie, Security Hub, and WAF) built in the right order from day one. We almost walked through the wrong door. A pre-launch architecture review with our engineers & AWS solution architects caught it in time. Here's how to have that conversation before it's too late.

Platform Engineering & DevOps

COM316 | Think Beyond DevOps: SRE, SOC, and Compliance in One AWS DevOps Agent (Chalk talk, 300)
Shaoyi Li ( @ShaoiLi ), Richard Fan ( @richardfan )
  • Mon, Nov 30, 2:30 PM · Wynn/Encore, Montrachet 1
  • Thu, Dec 3, 2:30 PM · Caesars Palace, Trevi
As AI speeds up releases, continuous security control assurance & detecting threats in millions of telemetry signals get harder. We transformed AWS DevOps Agent into one agent for SRE, SOC and compliance, effectively addressing AI-era cloud security challenges in production. With AWS Config Conformance Packs for compliance-as-code and security MCP for finding triage, it produces point-in-time compliance and well-architected assessments, & drives alarm-triggered agentic SOC investigations within a controlled agent space, automating responses and keeping human-in-the-loop on high-risk mitigations only.
COM328 | Autonomous Remediation with OpenTelemetry and DevOps Agent (Code talk, 300)
Julia Furst Morgado ( @juliafmorgado ), Jones Zachariah Noel ( @zachjonesnoel )
  • Mon, Nov 30, 12:00 PM · Caesars Forum
  • Wed, Dec 2, 4:30 PM · MGM Grand
You instrumented everything. You have metrics, logs, and traces. And when production breaks a human still has to piece it together, tab by tab, signal by signal, under pressure. That is not an observability problem, it’s an automation problem. We'll show you an architecture on AWS that closes it. OpenTelemetry captures signals across EKS, ECS, & Lambda. AWS DevOps Agent runs the investigation, finds the root cause, and prepares the fix. By the time a human looks at it, the pull request is ready. Come learn the triggers, investigation loop, & guardrails you need before letting an agent touch production.
COM324 | Pit Crew AI: Building Multi-Agent Cloud Operations on AWS (Code talk, 300)
Pablo Inchausti ( @pabloinchausti ), David Petrocelli ( @dpetro )
  • Wed, Dec 2, 3:00 PM · Caesars Forum
  • Thu, Dec 3, 11:30 AM · Caesars Palace
AI can generate code at racing speed, but production workloads still require continuous attention to security, reliability, cost, & operational risks. In this Code Talk, we build a multi-agent architecture on AWS inspired by a Formula 1 pit crew. Using Amazon Bedrock AgentCore, Strands Agents, MCP, GitHub, Jira, & Infrastructure as Code, we evolve from a single agent to a team of specialized agents for security, cost optimization, reliability, & observability. Attendees will leave with reusable architectures, orchestration patterns, & examples for cloud operations on AWS.

AI-Native Development

COM301 | Context Over Code: Making AI-Native Development Work at Scale (Breakout, 300)
Matt Houghton ( @mattdevdba1 ) · Thu, Dec 3, 3:00 PM · Caesars Forum
AI coding assistants generate code quickly, but that's not the problem most teams need solved. After rolling out an AI coding companion to 150+ engineers, we learned the real unlock was rethinking how we work, not how fast we type. This session covers our path from AI code companion to AI-Driven Development Lifecycle: building a Context Store with Bedrock Knowledge Bases and MCP to surface decades of institutional knowledge, implementing steering & verification patterns, and managing the organizational change that makes AI-native development stick. You'll leave with a framework for scaling AI beyond individual productivity.
COM401 | AI Didn't Wait for Security: Now What? (Breakout, 400)
Stav Ochakovski ( @stav ), Ran Isenberg ( @ranisenberg ) · Mon, Nov 30, 11:30 AM · Wynn/Encore, Encore Ballroom 5
AI didn't wait for your security review. Developers already ship code with agentic AI, coding assistants, & MCP servers without governance; blocking them lowers efficiency & leaves you falling behind. This builder-focused session dissects the architecture of a governed AI platform. You'll learn how to build a centralized MCP broker with Amazon Bedrock AgentCore Gateway & Runtime, and how agent requests gain identity and permissions. Finally, you'll see how AI SDLC frameworks bring consistency and higher-quality artifacts, with DevOps & security agents woven into key stages.
COM331 | AI Writes the Code. Who Reviews the Reviewer? (Lightning, 300)
Rossana Suarez ( @roxsross ) · Wed, Dec 2, 4:30 PM · Wynn/Encore
AI coding assistants now write, refactor, and even review code at scale but who's actually accountable when something breaks in production? This session explores how AI has reshaped the code review process: faster first-pass reviews, AI-generated suggestions, and agentic tools running directly in CI/CD pipelines. You'll leave with a practical framework for combining AI speed with human accountability in your own review process.
COM343 | Craft, Don't Vibe, Your MVP (Lightning, 300)
Matt Lewis ( @mattlewis ) · Mon, Nov 30, 5:30 PM · Venetian, Expo Builder Theater
Many people today use LLMs only for code generation. The best teams embrace AI across all stages of the SDLC. Rather than relying on vibe coding to make these decisions implicitly, agents can be used to evaluate architectural options, compare trade-offs and challenge assumptions before a single line of code is created. Learn from practical real world examples how AI is being used to accelerate the delivery of production-ready systems, rather than those that carry technical debt from day one.
COM342 | Renaissance Dad: building an agentic learning app for my 9-year-old with Kiro (Lightning, 300)
Sylvain Bruas ( @sylvainbruas ) · Mon, Nov 30, 6:30 PM · Venetian, Expo Builder Theater
Werner Vogels told us to become Renaissance Developers — curious polymaths who blend disciplines & own what they build. I took it literally: as a dad, I used Kiro and agentic AI to build a learning app that helps my 9-year-old practice math and school lessons. In this session, I'll share the full architecture: an OCR agent that imports exercises from photos of school notebooks, an adaptive agent that personalizes difficulty, and a coaching agent that guides instead of correcting. We'll deep-dive into multi-agent orchestration patterns, vision model integration, spec-driven development with Kiro, & the guardrails required when your end user is a child. Attendees will leave with reusable agentic patterns applicable far beyond education.

AI Agents in Production

COM308 | What Building an SRE Agent Taught Us About Agents at Scale (Breakout, 300)
Geethika Guruge ( @geethika ), Arshad Zackeriya ( @devopswithzack ), Dipin Thomas ( @dipinthomas ) · Wed, Dec 2, 2:30 PM · Caesars Forum
Anyone can wire up an agent, but operating one safely and cheaply across thousands of sessions is where developer craft begins. Grounded in a production SRE agent we built on Amazon Bedrock AgentCore & the Strands Agents SDK, this session dissects engineering around agentic models: session isolation, context management, and least privilege identity. What broke, and how we fixed it: an agent with too much access, context bloat that wrecked cost, a crash during a tool call that forced durable recovery.
COM319 | Stop Rolling Back AI Agents: Production Patterns That Work (Breakout, 300)
Kristine Armiyants ( @kristinearmiyants ), Nadia Reyhani ( @nafia1212 ), Gunnar Grosch ( @gunnargrosch ) · Wed, Dec 2, 4:30 PM · Caesars Palace, Roman I
The demo shipped. The agent went live. Then someone pulled it back. 74% of organizations that have deployed an AI agent have rolled it back or shut it down, & the rate rises to 81% among those reporting fully mature guardrails. We'll walk through the failure patterns behind rollbacks: agents that fail silently, prompts treated as safety boundaries, & monitoring blind to non-deterministic failures. You'll leave with the patterns that prevent them: infrastructure-level guardrails, blast radius containment, & observability built for probabilistic systems.
COM313 | Agentic commerce in production: multi-agent orchestration on AgentCore (Breakout, 300)
Ricardo Ceci ( @ricardoceci ), Matias Kreder ( @mkreder ) · Thu, Dec 3, 4:00 PM · MGM Grand, Grand 117
A single agent that both advises and sells gets fragile fast. This session shows a production multi-agent system on Amazon Bedrock AgentCore and Strands Agents, running for real Shopify merchants, where an orchestrator coordinates two specialists: a wellness advisor in a regulated domain and a commerce agent that owns decision and checkout. We walk the decisions that matter: orchestration topology, intent routing and handoffs, shared versus isolated memory across agents, where the action boundary & per-agent guardrails live, and the latency & token cost that coordination multiplies. You leave with an orchestration pattern, & the code to implement it directly in your online store.
COM318 | KYC in 90 Seconds: Real-Time Onboarding with AgentCore (Breakout, 300)
S Logesh ( @logesh182 ), Avinash Dalvi ( @aviboy2006 ), Poobalan P ( @poobalan1210 ) · Thu, Dec 3, 1:00 PM · Caesars Forum
Opening a bank account should take seconds, not days. This session shows how AI agents can speed up KYC while keeping the final decision under control, specialized agents that run on Agentcore Harness investigate identity, documents, fraud signals, and screening results, with checks running in parallel. AgentCore Gateway routes tool requests through Cedar policies before tools run, while deterministic rules decide whether a case moves forward or needs human review. The session follows a clean case, an ambiguous case, and a new risk event after onboarding. Attendees leave with a pattern for regulated agentic systems.
COM325 | The agent orchestration spectrum: Agents, Step Functions, or Lambda? (Code talk, 300)
Håkon Eriksen Drange ( @haakon )
  • Wed, Dec 2, 11:30 AM · Wynn/Encore
  • Thu, Dec 3, 3:00 PM · Caesars Forum
Orchestration mode matters more for end user experience than just model choice. In this live-coding session, I build the same GenAI workload three ways: Amazon Bedrock AgentCore, AWS Step Functions, and AWS Lambda durable functions - and show what each costs in latency, dollars, and debuggability on a production content platform. Leave with a reusable decision framework to match orchestration to your workload.
COM326 | Observable Agents with Lambda Durable Functions and OpenTelemetry (Code talk, 300)
Raphael Manke ( @raphaelmanke ), Gunnar Grosch ( @gunnargrosch ) · Mon, Nov 30, 3:00 PM · Caesars Forum
When AI agents run as long-lived, multi-invocation workflows on Lambda Durable Functions, standard observability breaks down. A single durable execution spanning five specialists producing independent telemetry, replays pollute logs with duplicates, and no tool can tell you which agents are failing across thousands of concurrent runs. We'll build a multi-agent purchasing coordinator with Lambda durable functions, instrument it with the OpenTelemetry plugin for one coherent trace across all invocations, and CloudWatch insights to answer fleet-wide queries with business context.
COM317 | Ship Reliable AI Agents with Strands Evaluation and AgentCore Evaluations (Code talk, 300)
Guille Ojeda ( @guilleojeda )
  • Tue, Dec 1, 1:00 PM · Wynn/Encore
  • Thu, Dec 3, 12:00 PM · Caesars Forum
Your AI agent works in demos, but can you trust it enough for production? In this 300-level Code Talk, we'll turn real-world agent failure modes into representative test cases, evaluate both answers and actions, & add risk-based quality gates into CI/CD. Using Strands Evaluation and Amazon Bedrock AgentCore Evaluations, you'll leave this session able to design and implement an evaluation pipeline tailored to your own agents, backed by a public reference implementation you can adapt.
COM312 | Giving Your AI Agent a Wallet: The Architecture of Spend Control (Chalk talk, 300)
Sebastian Marines ( @sebastianmarines ), David Victoria ( @vikomex )
  • Mon, Nov 30, 1:30 PM · Caesars Forum, Alliance 305
  • Thu, Dec 3, 2:30 PM · Caesars Palace, Octavius 11
AI agents can now hold a wallet and pay for APIs, data, and even other agents. The demo takes a few hours; making it safe to ship takes much longer. At the whiteboard, using Amazon Bedrock AgentCore Payments, x402 micropayments, & Cedar policy, we design an agent that spends under a hard budget, then break it together. We work through where guardrails belong, why session limits aren't enough, custodial versus non-custodial tradeoffs, and human-in-the-loop without human-in-the-way. Bring your edge cases. You leave with a reference architecture and a checklist for bounded agent autonomy.
COM321 | The Agent Harness You Own on AWS for Regulated Industries (Chalk talk, 300)
Jacky Chan ( @jhkchan )
  • Mon, Nov 30, 3:00 PM · Caesars Forum, Alliance 314
  • Thu, Dec 3, 2:30 PM · Wynn/Encore, Chopin 2
In regulated industries the model is easy; the governable, auditable, defensible agent is hard. AgentCore ships primitives, but the governance harness you own on top is the real work. We build the rails it does not give you: a fail-closed, blocklist-first permission gate where deny beats a broad allow; a hash-chained audit verifiable off-cloud; and a RACI gate binding every action to a named human. Each maps onto a managed AWS primitive — Identity, Guardrails, Evaluations, Lambda MicroVMs for isolation — naming where the primitive ends and your code begins. You leave with the rails.
COM311 | Your Users Aren't Waiting to Build AI Agents. Are You Ready? (Chalk talk, 300)
Rian Brooks-Kane ( @rianbk ), Ian Mckay ( @ianmckay )
  • Mon, Nov 30, 1:30 PM · Caesars Forum, Alliance 314
  • Tue, Dec 1, 11:30 AM · MGM Grand, Premier 311
The spreadsheet wizards and automation tinkerers in your organization are already reaching for low-code and visual builder tools to assemble AI agents on their own. Call it citizen agent building: either governance sprawl, or your fastest path to scale. This interactive chalk talk gives platform, security, and data leaders a reference pattern that puts low-code builders on top and Amazon Bedrock beneath, so the security, data residency, and governance enterprises need come from the foundation rather than from each agent. Amazon Bedrock makes governance possible; your architecture makes it the default.
COM341 | The AgentCore Governance Loop: Policy, Guardrails, Observability (Lightning, 300)
Harpreet Singh Siddhu ( @harpreetsiddhu ) · Mon, Nov 30, 5:00 PM · Venetian, Expo Builder Theater
Building an agent takes an afternoon. Operating it safely across a team takes architecture. In this session learn how to build a multi-agent reference platform on Amazon Bedrock AgentCore and walk the full governance loop including cataloging agents for discovery, enforcing what they can do with Cedar policy at the gateway, filtering what they can say with Guardrails, and tracing behavior through observability. You leave with a governance reference architecture you can adapt to any agentic workload, plus the lessons the documentation leaves out.
COM338 | From Demo to Production: The Craft Behind AI Agents at Scale (Lightning, 300)
Gerardo Arroyo Arce ( @codecr ), Diana Castro Cordero ( @diana22 ) · Wed, Dec 2, 9:30 AM · Wynn/Encore, Lightning Theater
Your agent reasons, calls tools, and demos beautifully. Then production happens. Using real AWS accounts and Amazon Bedrock AgentCore, I'll walk through deterministic guardrails that make destructive actions impossible (not just discouraged), spend limits an agent can't prompt-inject its way past, evaluation you can trust, and the IAM that silently breaks first. You'll leave with a production-readiness framework for agentic systems—and honest stories about everything that failed on the way there.
COM201 | Nine Seconds to Zero: Protecting AWS from AI Agent Disasters (Lightning, 200)
Jasdeep Singh Bhalla ( @jasdeepsinghbhalla ) · Wed, Dec 2, 3:30 PM · Wynn/Encore, Lightning Theater
In April 2026, an AI coding agent deleted a company's entire production database in nine seconds, without being asked. The agent was confidently wrong and had the permissions to act on it. In this session, we recreate that incident live using Amazon Bedrock, walk through every failure in the chain, then rebuild the architecture with six hardening layers: Bedrock Guardrails, human approval gates, least-privilege IAM, deletion protection, isolated backups, and real-time monitoring. Walk away with a concrete checklist for designing AI-assisted systems that plan for when the AI is wrong, not just when it performs perfectly.
COM346 | Architecting Enterprise Agents on Bedrock (Lightning, 300)
Namrata Shah ( @namratashah ), Elliott Cordo ( @elliottcordo ) · Wed, Dec 2, 5:00 PM · Wynn/Encore, Lightning Theater
Enterprise agents on Amazon Bedrock are complex distributed systems, combining models, tools, memory, and orchestration into something that has to behave predictably in production. In this session, I walk through practical architecture patterns for building enterprise-grade agents: multi-agent coordination, tool routing, session management, and hybrid workflows using AWS Step Functions and AWS Lambda. Walk away with concrete patterns you can bring back to your own agent architecture.
COM339 | Cocktails to Code: Building AI Bartenders with Amazon Bedrock (Lightning, 300)
Jimmy Dahlqvist ( @jimmydqv ) · Tue, Dec 1, 5:00 PM · Venetian, Expo Builder Theater
Plain-language AI orders fail if models hallucinate items that are out of stock. I built a reliable AI bartender grounded strictly in live inventory data. This session covers how an AI agent on Amazon Bedrock uses Strands Agents, AgentCore, and MCP tools to fetch real-time menus. We will dive into the serverless architecture, see what breaks during live text streaming, and evaluate the cost versus performance tradeoffs of production grounding. You will leave with a reusable pattern and the repository to build reliable, data-driven agents.

Generative AI, Data & Cost

COM305 | Why Enterprise RAG Fails: Document AI and Ontology Patterns on AWS (Breakout, 300)
Ayyanar Jeyakrishnan ( @ayyanar ), Bhuvaneswari Subramani ( @bhuvaneswari ), Vivek Raja P S ( @vivekraja ) · Thu, Dec 3, 9:00 AM · Wynn/Encore, Purple Theater
Most teams blame embeddings, vector databases, or foundation models when RAG systems underperform. In practice, failures often originate earlier in the pipeline where business meaning is lost during document ingestion. This session explores production Document AI patterns on AWS, including layout-aware extraction, ontology-driven classification, semantic chunking, metadata enrichment, and retrieval optimization. Attendees will learn practical architecture patterns, implementation tradeoffs, and a framework for building scalable, trustworthy RAG systems for enterprise workloads.
COM322 | Eval-Driven LLM Routing: Cut Costs Without Losing Accuracy (Chalk talk, 300)
Luca Bianchi ( @aletheia ), Olivier Leplus · Wed, Dec 2, 4:30 PM · Caesars Forum, Alliance 305
This session shows how to build an eval-driven LLM routing system: harvest a production eval dataset, define accuracy KPIs using LLM-as-judge, set a non-negotiable accuracy floor as the practitioner's decision, benchmark candidate models, and generate a routing policy optimized for cost within that floor. Result: approximately 80% cost reduction at maintained accuracy. A reference architecture you can start implementing immediately.
COM347 | Beyond Prompts: Crafting Production GenAI from Scratch (Lightning, 300)
Steve Leung ( @steveleung ) · Wed, Dec 2, 9:00 AM · Wynn/Encore, Lightning Theater
Building reliable GenAI applications requires deliberate systems engineering, not just prompt engineering. After collaborating with AWS Professional Services on an enterprise client, I learned how to build production-grade web apps the official AWS way.
Bypassing basic buzzwords, we map a compound serverless conversational assistant engineered from scratch using AWS CDK, CloudFront, Cognito, API Gateway, and Amazon Bedrock.
Examine how to handle real-world token streaming failures, manage state in DynamoDB, and implement strict Infrastructure as Code. Leave with complete source code and a concrete approach to balancing model performance against infrastructure cost in highly concurrent production environments.
COM327 | Answering CX questions 10x faster with Amazon Connect Customer and Amazon Quick (Code talk, 300)
Kobe Xu ( @kobexu ) · Thu, Dec 3, 4:00 PM · MGM Grand, Boulevard 159
Amazon Connect Customer generates a deep set of signals from every customer interaction. Sentiment, intent, conversation analytics, resolution outcomes. It's all there & Connect Customer gives you built-in tools to explore it. But we wanted to go further. We wanted anyone in the organization to ask questions about customer interactions in plain language and get answers instantly. So we paired Connect Customer's analytics data lake with Amazon Quick. In this session, we walk through how we did it, including how we handled cross-account Lake Formation access, and show how our broader team now asks questions and reshapes queries on the fly. 10x faster from question to action.

Learn From the Community, All Week Long

These speakers have done the work. They've scaled serverless workloads by orders of magnitude, run GPU inference across cloud & bare metal, recovered from ransomware, passed audits their scanners missed, & kept AI agents in production when others rolled theirs back. Few places offer this much hard-won experience across this many topics: compute, containers, resilience, security, platform engineering, & generative AI.
re:Invent lasts a week, but the community behind these talks is around all year.
Come learn from the builders. Stay for the community.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article