Loading…
AI Agent Development · Seattle
Most agent projects in Seattle live inside an AWS or Microsoft environment, and most fail at the same place: production operations. The prototype works in a notebook against three test cases. The production agent has to survive Bedrock rate limits, Azure OpenAI throttles, idempotency requirements when writing to Stripe or Salesforce, retry policies that do not cascade into a 503 storm, and an observability story that lets an SRE answer questions at 2am.
We build agents for AWS-native and Azure-native teams in the metro. Bedrock Agents when the workflow fits, LangGraph on Bedrock Runtime when it does not. Semantic Kernel and Azure AI Foundry for Microsoft stacks. OpenTelemetry spans into Datadog or Honeycomb. SQS or EventBridge between the agent and any downstream side effect.
Fixed-scope projects are quoted on scope. Discovery first, fixed quote after, four to eight weeks of build, weekly demos on Pacific time.
Tell us about the production agent.
On AWS, the decision tree usually starts with Bedrock Agents vs custom orchestration. Bedrock Agents is the managed path: you define action groups backed by Lambda, point at a Bedrock Knowledge Base for retrieval, and the runtime handles the loop. It fits when the workflow maps cleanly to action groups and you want to ship without owning the orchestrator. It stops fitting when you need conditional state transitions, parallel tool calls in a fan-out pattern, or a memory backend that is not Bedrock's.
When the workflow exceeds what Bedrock Agents handles, we ship LangGraph on ECS Fargate or Lambda, with Bedrock Runtime as the model layer for Claude Sonnet 5 or Llama 4, and Bedrock Knowledge Bases retained for retrieval if it still fits. The graph is checkpointed in DynamoDB so a Lambda timeout does not throw away the trajectory. State transitions are explicit. A Step Functions state machine wraps the long-running agent runs where each step might take minutes.
Idempotency on every downstream side effect is non-negotiable. Every Stripe charge, every Salesforce write, every Zendesk ticket update gets an idempotency key derived from the agent run ID plus the action hash. A retry never double-writes. We use SQS or EventBridge between the agent and the side-effect worker so a spike in agent traffic does not cascade into a downstream outage.
On Azure, the equivalent decision is Azure AI Foundry vs Semantic Kernel vs custom. Azure AI Foundry gives you a managed agent runtime with built-in tool calling against GPT-5 or Claude on Azure, plus tracing through Application Insights. It fits Microsoft-first organizations that want the agent reachable from Copilot Studio, Teams, and Power Automate without extra glue.
Semantic Kernel works when you need more control of the planner, want to mix .NET and Python services, or have an existing Azure Functions or Container Apps stack you want the agent to live inside. We have shipped Semantic Kernel agents that bridge Microsoft Graph (Outlook, SharePoint, Teams) with external systems through MCP tools.
For observability across either cloud, we instrument with OpenTelemetry spans into Datadog APM, Honeycomb, or Grafana Tempo. Every LLM call is a span with model, token counts, latency, cost. Every tool call is a span with tool name, status, downstream latency. This is the difference between an agent you can debug at 2am and one you cannot.
Six engineering components that show up in every production agent we ship for cloud-native, e-commerce, and marketplace teams.
LangGraph on ECS Fargate or Lambda for AWS shops; Semantic Kernel on Azure Container Apps for Microsoft shops. Checkpointed state in DynamoDB or Cosmos DB so a runtime restart resumes from the last completed node.
Model layer pinned by version (Claude Sonnet 5, GPT-5, Llama 4 70B). Cross-region inference profiles on Bedrock. Provisioned throughput where the workload justifies it. Cost dashboards by feature.
Token-bucket throttle that respects Bedrock TPM and RPM headers and Azure OpenAI quota responses. Exponential backoff with jitter, circuit breakers, fallback to a cheaper model when primary is throttled. SQS or EventBridge between agent and downstream.
Idempotency keys derived from agent run ID plus action hash. Stripe charges, Salesforce writes, Zendesk updates, Snowflake inserts. A retry never double-writes. Failure cases route to a DLQ with the reasoning attached.
Every LLM call is a span with model, token counts, latency, cost. Every tool call is a span with tool name, status, downstream latency. Exports to Datadog APM, Honeycomb, or Grafana Tempo. Searchable by trajectory length and override flag.
Braintrust or Promptfoo against a labeled set, gating CI. Task completion rate, hallucination rate, tool-call accuracy, latency p50 and p95. Regression suite runs on every prompt or model change.
Seattle is where the cloud and the agent industries overlap most tightly. Amazon ships Bedrock from South Lake Union, Microsoft ships Azure AI Foundry from Redmond, and Auth0 (now Okta), Zillow, Expedia, Redfin, and Tableau all run sizable engineering organizations that consume those services. The teams here have deep cloud operating experience but are still working out the agent-specific operating story: rate limits across thousands of concurrent runs, idempotency at marketplace scale, and observability that supports an on-call SRE rather than a prototype developer.
The e-commerce and marketplace shape of the local economy is part of why production agent ops matters here more than in cities where most agents are prototypes. A marketplace customer-support agent at Expedia or Redfin scale will see millions of runs a month. A pricing or inventory agent at an Amazon FBA shop or a Shopify Plus store on Mercer Island sees thousands of write actions a day to inventory and listing systems. The same code that worked at 100 runs a day becomes a 5xx storm at 10,000 without the idempotency and rate-limit layers in place.
We work remotely with Seattle clients on Pacific time. Discovery calls fit between 9am and 5pm PT, demos happen weekly, and the codebase ships to a repository you own.
Industries where we see strongest fit: cloud-native B2B SaaS, e-commerce and marketplace ops, travel and booking platforms, customer support automation, and engineering teams modernizing existing automation onto Bedrock or Azure AI Foundry.
Describe the agent, the cloud you live on, and the production constraints (volume, SLOs, downstream systems). We'll reply within one business day with a rough scope, a recommended architecture, and a price range.