Loading…
AI Agent Development · Minneapolis
The Twin Cities Fortune 500 cluster runs operations at a scale where even a 1 percent automation lift is measurable money. UnitedHealth Optum processes roughly 1.6 billion claims a year. Target moves merchandise across 1,956 stores and 38 regional distribution centers. Cargill trades commodities across dozens of desks. 3M and Medtronic each handle quality and complaint events under FDA 21 CFR Part 820. An agent that handles tier-1 work at $0.10 to $0.30 per run versus $4 to $8 per run for a human pays back quickly.
We build agents for these environments without the production-ops gaps that kill prototype agents in week one. LangGraph on ECS Fargate or EKS, idempotency keys on every side effect, token-bucket throttles in front of every external system, OpenTelemetry spans into Datadog or Honeycomb. Model layer through Bedrock Claude or Azure OpenAI under your existing enterprise BAA or DPA.
We scope projects fixed and flat, with pricing tailored to the workflow and volume. Discovery includes a production-readiness review against your existing SRE and audit standards.
Tell us about the workflow and the volume.
An Optum prior-authorization analyst at $90,000 fully loaded handles roughly 60 cases per day. A tier-1 case that an agent can complete end-to-end (eligibility check, policy lookup, criteria match, automated notification) takes the agent 90 to 180 seconds and roughly $0.10 to $0.30 in model and infrastructure cost. The same case takes a human 8 to 12 minutes including system context switches. At 60 cases a day the analyst burns three to five hours on tier-1 work that an agent can handle. Over a 200-analyst department that is 700 to 1,000 hours a day of recoverable capacity.
The same math applies at Target merchandising, where supplier-portal exception handling consumes hours per buyer per day. At Cargill middle-office, where settlement reconciliation and trade-break investigation pull analysts out of value-adding work. At 3M and Medtronic quality operations, where initial complaint classification and triage adds up to substantial engineering time.
The architecture has to match the operational reality. Agent runtime on ECS Fargate or EKS with horizontal autoscaling driven by queue depth, not CPU. Bedrock Claude Haiku 4.5 for routine work, Sonnet 4.5 for complex cases, with provisioned throughput for predictable capacity. State checkpointed in DynamoDB so a runtime restart resumes from the last completed node.
Idempotency is non-negotiable. Every claim status update, supplier rejection, planogram change, or settlement adjustment carries an idempotency key derived from the business key and action hash. A retry never double-writes. SQS or EventBridge sits between the agent and any downstream side effect so an agent traffic spike does not cascade into a payer-portal outage or a SAP-side meltdown.
Observability is OpenTelemetry into Datadog, Honeycomb, or Splunk depending on what your SRE team already operates. Every LLM call is a span, every tool call is a span, every agent run is a parent span with trajectory length and override flags as searchable attributes.
Six engineering components that show up in every Fortune 500 agent we ship in the metro.
LangGraph on ECS Fargate or EKS with checkpointed state in DynamoDB. Horizontal autoscaling driven by SQS queue depth. Survives runtime restarts without losing trajectory.
Bedrock Claude Haiku 4.5 for tier-1 work at low cost, Sonnet 4.5 for complex cases. Provisioned throughput for predictable capacity. Cost dashboards per workflow.
Claim updates, supplier rejections, planogram changes, settlement adjustments all keyed by business ID plus action hash. A retry never double-writes. DLQ routing for failure cases with full reasoning trace.
Configurable human-in-the-loop for FDA-regulated complaint handling, trading writes, payer-portal submissions. Agent proposes, human confirms. Configuration per workflow rather than hard-coded.
Datadog, Honeycomb, or Splunk integration. Spans for every LLM call and tool call. Searchable by trajectory length, business key, and override flag. SRE can debug at 2am from a trace.
Braintrust or Promptfoo against a labeled real-case set. Task completion rate, hallucination rate, tool-call accuracy, regression suite in CI. Doubles as Operational Qualification documentation for regulated workflows.
The Twin Cities Fortune 500 concentration is the densest in the country per capita. UnitedHealth Group in Minnetonka, Target downtown, 3M in Maplewood, Cargill in Wayzata, Best Buy in Richfield, US Bank in the IDS Center, plus Medtronic, Ecolab, Securian, Allianz Life, and Polaris. Each of these has back-office and operational workflows where automation compounds across thousands of analysts, buyers, traders, or quality engineers.
The procurement culture in the metro is conservative for good reason. Healthcare-payer workflows are CMS-anchored. Medical-device complaint handling is FDA-audited. Banking and asset-management workflows are OCC and SEC examined. Retail supplier workflows are GS1 and SOX governed. An agent that ships into one of these environments has to come with an audit story and an SRE runbook that matches the existing operational standard.
We work remotely with Twin Cities clients on Central time. Discovery includes a production-readiness review against your existing SRE, audit, and change-management standards. Code ships to repos you own with full IaC.
Industries where we see strongest fit: healthcare-payer claim and prior-auth ops, retail supplier and merchandising ops, medical-device quality and complaint handling, agribusiness middle-office and settlement, banking back-office, and manufacturing quality.
Describe the workflow, the volume, the regulatory frame, and the systems the agent has to touch. We'll reply within one business day with a rough scope and a price range.