AI-Powered Transformation of AWS Support for Mission-Critical Workloads

whoop · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Cloud support and mission-critical workload operations

AWS Support needed to move from reactive incident firefighting to proactive, context-aware assistance for increasingly complex cloud workloads, while WHOOP needed dependable launch support for traffic spikes around new hardware.

Constraints:

Architecture

AWS Support built an AI-augmented support platform around customer workload context, Bedrock agents, Amazon Connect channels, MCP-exposed tools, structured SOPs, orchestration, and human experts. Customer channels such as console, chat, and Slack enter a central layer that selects the right agent or expert. A context service hydrates workload, application, and business-specific details for both humans and AI. Unstructured runbooks are converted into deterministic workflows, tools, and agents, validated with integration-style evaluations. WHOOP consumed this capability through Unified Operations, Countdown Premium, proactive risk assessment, dedicated experts, and incident response during a large product launch.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  support-channel-layer["support-channel-layer"] --> agent-orchestration-layer["agent-orchestration-layer"]
  support-channel-layer["support-channel-layer"] --> amazon-connect["amazon-connect"]
  agent-orchestration-layer["agent-orchestration-layer"] --> bedrock-agentcore["bedrock-agentcore"]
  agent-orchestration-layer["agent-orchestration-layer"] --> context-service["context-service"]
  agent-orchestration-layer["agent-orchestration-layer"] --> human-support-experts["human-support-experts"]
  context-service["context-service"] --> customer-workload-context["customer-workload-context"]
  context-service["context-service"] --> human-support-experts["human-support-experts"]
  bedrock-agentcore["bedrock-agentcore"] --> mcp-tool-and-sop-layer["mcp-tool-and-sop-layer"]
  bedrock-agentcore["bedrock-agentcore"] --> automated-reasoning-checks["automated-reasoning-checks"]
  mcp-tool-and-sop-layer["mcp-tool-and-sop-layer"] --> authoring-and-evaluation-platform["authoring-and-evaluation-platform"]
  mcp-tool-and-sop-layer["mcp-tool-and-sop-layer"] --> customer-workload-context["customer-workload-context"]
  automated-reasoning-checks["automated-reasoning-checks"] --> human-support-experts["human-support-experts"]
  amazon-connect["amazon-connect"] --> human-support-experts["human-support-experts"]
  whoop-unified-operations["whoop-unified-operations"] --> human-support-experts["human-support-experts"]
  whoop-unified-operations["whoop-unified-operations"] --> customer-workload-context["customer-workload-context"]
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
support-channel-layer Receives customer engagements from console, chat, Slack, Connect, and support programs. agent-orchestration-layer, amazon-connect
agent-orchestration-layer Selects the right agent or multiple agents for a support issue and reconciles their outputs. support-channel-layer, bedrock-agentcore, context-service, human-support-experts
context-service Hydrates customer workload, application, architecture, and business context for agents and humans. agent-orchestration-layer, customer-workload-context, human-support-experts
bedrock-agentcore Provides agent runtimes and gateways that expose tools in MCP form. agent-orchestration-layer, mcp-tool-and-sop-layer, automated-reasoning-checks
mcp-tool-and-sop-layer Structured workflows, deterministic tools, and specialized agents derived from support SOPs and runbooks. bedrock-agentcore, authoring-and-evaluation-platform, customer-workload-context
authoring-and-evaluation-platform Converts natural-language workflows into model-optimized tools with integration tests and evaluation cases. mcp-tool-and-sop-layer
automated-reasoning-checks Governance layer for safe and accurate responses where non-overlapping rules can be checked near-deterministically. bedrock-agentcore, human-support-experts
amazon-connect Customer-interaction platform with summarization, translation, and contact-center AI features. support-channel-layer, human-support-experts
human-support-experts TAMs, engineers, incident managers, and service experts who use agent diagnostics and customer context. agent-orchestration-layer, context-service, automated-reasoning-checks, amazon-connect
customer-workload-context Application-specific architecture, runbooks, failure modes, business logic, and launch-readiness information. context-service, mcp-tool-and-sop-layer
whoop-unified-operations WHOOP's use of Countdown Premium, Unified Operations, risk assessments, launch-room support, and critical incident response. human-support-experts, customer-workload-context

Data flow

Support interactions arrive from console, chat, Slack, or customer programs. The central orchestration layer selects a general or specialized agent and may fan out to multiple agents. The context service supplies workload-specific context. Bedrock AgentCore runtimes and gateways expose tools and SOPs through MCP. Agents and expert humans use those diagnostic outputs to reproduce issues, reason from symptoms to root cause, recommend mitigations, or surface safe responses through Connect and support workflows.

4 supporting quotes
“AWS support today, that includes in the console, chat, Slack, it already exists. And we have these coming into a central layer.”transcript, stated
“this orchestration layer, what it does is it selects the right agent for the job.”transcript, stated
“we also have built a context service, which ensures that our agents can build this context needed to solve customer issues in a personalized way”transcript, stated
“Agent Core has been pretty fantastic in allowing us to have runtimes and have gateways to just convert our tools into MCP.”transcript, stated

Why this architecture

AWS chose a context-first, agent-plus-human architecture because generic support chatbots cannot reliably solve complex, mission-critical cloud incidents. Structured SOPs and MCP tools make agent execution repeatable, while orchestration supports specialized agents instead of one giant generalist. Human experts stay central for complex issues, but they receive richer diagnostic outputs and workload context. WHOOP used the same proactive context model for launch readiness, making support staff effectively an extension of its engineering team.

4 supporting quotes
“one thing we learned was context is key, and the AI is as good as the data behind it.”transcript, stated
“We wanna detect issues before they impact your production.”transcript, stated
“to really get it to run as an agent or for an agentic execution to do this reliably as a tool in a way that we can trust, there is more of a structured format.”transcript, stated
“they understood our systems already, they were able to help us resolve the issues faster as well.”transcript, stated

Models used: Amazon Bedrock · Bedrock AgentCore · Amazon Connect · MCP · Graph RAG · LoRA fine-tuning · automated reasoning checks

Scale

Alternatives considered (roads not taken)

Reactive firefighting support

Rejected because: It is heroic but not scalable for complex, distributed customer workloads and recurring incident patterns.

“cloud support over the last few years have been like firefighters, right? An event happens, you contact us, we help you get over that event.”transcript, stated
“It's not something that's scalable. A problem sparks, an alarm goes off, your team investigates, you contact us, we tell you where there”transcript, stated

One general support agent

Rejected because: A single agent works for general guidance but not for complex support situations needing specialized agents and reconciliation.

“if we had a single agent, which is how we started, for example, for general guidance, then it works fine.”transcript, stated
“when there is an agentic system involving multiple agents, there is an orchestration layer that's needed.”transcript, stated

Manual SOP/runbook execution

Rejected because: Manual and unstructured runbooks do not scale across hundreds of services, so AWS converts them into deterministic workflows, tools, and agents.

“we invested in moving these unstructured SOPs that we already had into structured SOPs using AI.”transcript, stated
“that approach is hard to scale and so we also started looking at alternative approaches of scanning our corpuses of data and so this included our”transcript, stated

Plain RAG over support knowledge

Rejected because: AWS found graph-based retrieval richer for troubleshooting paths from symptoms to root cause.

“moving from a rag to a graph rag was pretty helpful. So instead of just retrieving knowledge in a specific way through chunks, having the graph-based rag led to more richer knowledge retrieval.”transcript, stated

Patterns

multi-agent-orchestration (agent architecture)

AWS uses an orchestration layer to route work to one or more specialized support agents and reconcile richer responses.

“there are going to be multiple agents that are needed to solve a problem.”transcript, stated
“this orchestration layer, what it does is it selects the right agent for the job.”transcript, stated

tool-use-orchestration (agent architecture)

Support SOPs and diagnostic automations are exposed as MCP tools that agents can call during troubleshooting.

“the way that this is being surfaced across agents, these tools and these SOPs is via MCP.”transcript, stated

rag-retrieval (retrieval)

AWS moved from chunk-based RAG toward graph RAG for richer symptom-to-root-cause reasoning over support knowledge.

“having the graph-based rag led to more richer knowledge retrieval.”transcript, stated

eval-driven-development (evaluation)

Converted SOPs include integration-style tests that verify agents can reproduce and solve issues in AWS accounts.

“actually implementing that in that account and if it was able to solve the problem, then that's a pass, right?”transcript, stated
“it's a very complete evaluation in that scenario and that's what we're doing for this SOP approach.”transcript, stated

guardrails-io-filtering (safety)

AWS uses automated reasoning checks and governance controls to detect unsafe or out-of-bounds responses.

“automated reasoning checks have been pretty useful.”transcript, stated
“near deterministically being able to tell us if something is going beyond some safety check.”transcript, stated

Tradeoffs

Keep humans in control for complex operational support.

Chose AI augmentation plus expert humans over Fully autonomous support for all incidents.

Rationale: AWS distinguishes assistance, augmentation, and autonomous AI, allowing automation based on confidence while still using experts for complex issues.

“So human is taking some actions and the human is still in control, but AI is kind of augmenting.”transcript, stated
“then the third tier is where you have autonomous AI, where AI is acting always in your environment”transcript, stated

Invest in context and evaluations as durable assets.

Chose Context service plus evaluation suites over One-off model/framework tuning.

Rationale: Models and frameworks change, but workload context and tests continue to improve future support systems.

“technologies change, models are going to change, frameworks will change, but the evaluations and the context that you have are really gonna be durable.”transcript, stated

Use proactive launch support instead of waiting for production incidents.

Chose Risk assessment, load testing, designated experts, and launch room support over Reactive post-failure escalation.

Rationale: WHOOP's launch preparation surfaced risks before launch and put context-aware experts in place during the traffic spike.

“one of the earliest things we did was we engaged with AWS Countdown Premium, which is a part of Unified Operations, and we did a risk assessment exercise.”transcript, stated
“We had an AWS room designated where we had our TAMs and they were on the call with multiple experts from Unified Operations the whole day”transcript, stated

Gotchas (what bit them)

Generic AI guidance can become chaos without customer context.

Root cause: Support issues depend on each customer's workload, application architecture, business logic, and failure modes.

Resolution: Capture, hydrate, and feed workload context into both human expert workflows and AI agents.

“if we did not add a context around you, your workloads, your application-specific details, sometimes it could lead to chaos versus intelligence.”transcript, stated
“as we onboard your workloads, we capture that context, we hydrate it, and as we provide that context to humans, we also feed that into AI.”transcript, stated

Converting expert SOPs into reliable tools is heavy work.

Root cause: Unstructured runbooks contain commands, tribal knowledge, and incomplete assumptions that agents cannot safely execute as-is.

Resolution: Create structured SOPs, deterministic workflows, integration tests, and an authoring platform that experts validate.

“There needs to be completeness as well.”transcript, stated
“We really need our experts to come in, validate each approach. It's time consuming, it's heavy”transcript, stated

Too many MCP tools can hurt latency and accuracy.

Root cause: The model has to choose among hundreds of tools and options across AWS support scenarios.

Resolution: Use fine-tuning, LoRA, and retrieval improvements to help tool selection and targeted reasoning.

“Because the model has to go figure out which tool to use and if it's hundreds of tools, which it is in the case of AWS, it can lead to latency and it can lead to accuracy.”transcript, stated
“using fine tuning and also RFD in this approach. In case you don't know where to start with fine tuning, LoRa has been pretty great.”transcript, stated

Evaluation

AWS evaluates support automation through SOP integration tests, Bedrock AgentCore evaluations, automated reasoning checks, prevention metrics, automation outcomes, mean time to resolution, and customer outcomes such as WHOOP availability and response-time changes.

Metrics: SOP pass/fail in AWS accounts · prevention · full automation · mean time to resolution · critical case response time · availability · downtime reduction · load-test multipliers

4 supporting quotes
“have the agent figure out if it can solve the problem and once it comes up with a resolution, actually implementing that in that account”transcript, stated
“we also launched Bedrock Agent Core evaluations, which also allows easy evaluations out of the box.”transcript, stated
“the success metrics that we looked at were prevention, full automation,”transcript, stated
“critical case response time in AWS going from around just under eight minutes to under two and a half minutes.”transcript, stated

Outcomes

The support platform addresses a 15,000-person global AWS Support organization, 250-plus AWS services, and customer launches with multipliers like 10x mobile traffic and 200x e-commerce traffic.

4 supporting quotes
“the entire month of May was saw 100% availability”transcript, stated
“we were able to reduce that downtime 70% and achieve 99.95% availability in Q3”transcript, stated
“critical case response time in AWS going from around just under eight minutes to under two and a half minutes.”transcript, stated
“we feel confident going into the Christmas season and having people unbox their Christmas presents and start their New Year's resolution”transcript, stated