How Intuit built an agentic disaster recovery assistant with Amazon Bedrock
Intuit built "EWOK Agent," an AI-powered disaster recovery assistant using Amazon Bedrock, to automate decision-making in failover scenarios that previously required tribal knowledge from senior engineers The system separates reasoning from execution: foundation models decide what to do, while the deterministic EWOK (Ecosystem Wide Orchestrator Kit) layer executes how, maintaining a crisp boundary between the two EWOK already reduced recovery times from several hours to ~20 minutes by standardiz
Analysis
TL;DR
- Intuit built "EWOK Agent," an AI-powered disaster recovery assistant using Amazon Bedrock, to automate decision-making in failover scenarios that previously required tribal knowledge from senior engineers
- The system separates reasoning from execution: foundation models decide what to do, while the deterministic EWOK (Ecosystem Wide Orchestrator Kit) layer executes how, maintaining a crisp boundary between the two
- EWOK already reduced recovery times from several hours to ~20 minutes by standardizing failover execution across compute, databases, networking, caches, and async workloads via YAML-declared recovery intent
- Amazon Bedrock provides model flexibility (hundreds of FMs via a single API), built-in Guardrails for safety, encryption in transit/at rest, and no data training usage — critical for production financial systems
- The agent is delivered as a plugin installable from Intuit's Engineering portal or IDE, and the underlying design pattern (typed skills, thin Bedrock layer, bounded agentic loop over a deterministic executor) is reusable for any system with authenticated, auditable APIs
Why It Matters
This represents a practical, production-grade example of agentic AI applied to critical infrastructure operations — not just a demo, but a system actively used by Intuit teams for eight months to run failovers for products serving millions of users (TurboTax, QuickBooks, Mailchimp, Credit Karma). It demonstrates how to responsibly combine LLM reasoning with deterministic execution in high-stakes environments where safety, auditability, and policy compliance are non-negotiable.
Technical Details
- Architecture pattern: A thin Amazon Bedrock layer sits above the EWOK execution engine, connecting foundation models to "skills" (typed, compiled representations of failover knowledge). The agentic loop processes natural-language requests (e.g., "failover payments-gateway in production") into validated, policy-aligned recovery executions.
- Core design principle: "The model decides what to do, and the EWOK Agent deterministically executes how." This boundary ensures that LLMs handle ambiguous decision-making (workflow selection, exception handling, change-freeze override procedures) while the underlying system guarantees safe, auditable infrastructure actions.
- Amazon Bedrock integration: Uses the Bedrock Converse API with tool use (function calling) for skill-to-tool compilation. Supports model evaluation and switching without rearchitecting. IAM actions include
bedrock:Converse,bedrock:ConverseStream,bedrock:InvokeModel, andbedrock:ApplyGuardrail. - Security and compliance: Bedrock Guardrails provide built-in safety controls. Data is encrypted in transit and at rest, and is not used to train foundation models — essential for a system operating on production financial infrastructure.
- Delivery and stack: Delivered as a plugin for Intuit's Engineering portal and IDEs. Built with Python, AWS SDK for Python (Boto3), and langchain-aws libraries. The pattern is framework-agnostic and applicable to any system exposing authenticated, auditable APIs.
Industry Insight
- Agentic AI for infrastructure ops is viable at scale: Intuit's eight-month production deployment proves that AI agents can safely augment critical SRE workflows when the reasoning-execution boundary is strictly enforced — a pattern other enterprises should replicate for incident response, change management, and compliance-heavy operations.
- Model flexibility without lock-in is a strategic advantage: By using Bedrock's multi-model API, Intuit can continuously evaluate and swap foundation models as capabilities evolve, keeping the agent current without engineering overhead — a model that reduces long-term AI infrastructure costs.
- The "skills as compiled tools" pattern is broadly reusable: Encoding domain knowledge as typed skills that compile into function-calling tools, then wrapping them in a bounded agentic loop, provides a template for building reliable AI assistants in any regulated or safety-critical domain beyond disaster recovery.
Disclaimer: The above content is generated by AI and is for reference only.