
AWS Enterprise AI Agent Architecture: 3 Patterns, 3 Traps
Cloud Platform EngineeringAWS’s own architects say a common production setting is 3 to 5 loop iterations against a total token budget, and that an agent without that ceiling “can spend tokens indefinitely and still get nowhere.” That warning comes from a new AWS reference architecture for enterprise AI agents. For platform teams putting agents near real systems, the useful part is not the five-layer diagram. It is the short list of patterns and traps AWS Professional Services says it has seen in production.
The Framework
AWS’s post describes five layers. An API gateway handles identity and policy. A five-phase runtime loops through Understand, Plan, Retrieve, Reason, and Act and Observe. A knowledge and state tier provides memory, a platform-services layer handles models and tools, and a trust and control framework runs across all of them. The authors say they aren’t proposing new agent theory: the phases combine the robotics sense-plan-act loop, ReAct prompting and Reflexion-style self-critique under one vocabulary, so each phase has a clear place for governance.
Every output passes four gates: guardrails, evaluations, confidence scoring and human approval. AWS maps them to the NIST AI Risk Management Framework and the Well-Architected Generative AI Lens security guidance. Outcomes are tiered by risk. A plain response needs only guardrails. A write needs a policy gate and logging. Triggering a workflow needs policy plus human approval.
1. Pattern: Graduated Autonomy
Start agents read-only and widen write access step by step as they earn trust, using feature flags or IAM role escalation per user, team or use case. Identity makes this work: the gateway attaches the caller’s role, permissions and session to each request, and that context follows every tool call and data lookup.
2. Pattern: Match the Model to the Phase
Small, fast models handle intent classification and routing. Larger models are reserved for planning and reasoning, and specialized models run the evaluations. AWS says teams that route this way usually cut cost while keeping quality, but adds that savings depend on the workload and should be benchmarked first.
3. Pattern: Structured Tool Interfaces
Give every tool an OpenAPI schema with input and output types, error codes, rate limits and declared side effects. The planner uses it to choose tools, and the observe step checks results against what the tool promised.
The Traps
AWS names three anti-patterns. A monolithic agent that does everything should be split into specialized agents under a supervisor. Unbounded loops need hard caps on iterations, time and tokens, plus an escalation path. Otherwise latency rises, downstream APIs absorb redundant load, and spend climbs with nothing done. Trust-later architecture, adding guardrails and observability after launch, costs more and brings more defects than building them in from the start.
The post backs this with customer results. AWS cites Druva’s DruAI, built on Amazon Bedrock AgentCore, which resolves 68% of customer issues through agent workflows, and Cox Automotive, which moved 17 agentic solutions into production in under a year.
Where It Doesn’t Apply
AWS calls its service mapping a set of starting points, not a fixed bill of materials. That mapping covers API Gateway with Cognito, Bedrock Agents and Guardrails, Knowledge Bases on OpenSearch Serverless, DynamoDB for state, and X-Ray, CloudWatch and CloudTrail for observability. Scale, data residency, latency and cost will change some of those choices. Compliance stays with the customer under the shared responsibility model. Stack Archive’s earlier AgentCore architecture analysis covers the managed runtime alternative.
The So What
Platform teams preparing an agent for production should set a hard loop ceiling and a per-session token budget, and give every tool an OpenAPI schema with declared side effects, before the first pilot. Launch the agent read-only, and require human approval for any action that triggers a workflow until its evaluation and confidence-score data justify widening access.
Source: AWS Architecture Blog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →