AI agents need handoff architecture

On 5 August 2026, AWS published two production examples for AI agents: LendingTree runs a multi-part mortgage assistant on Amazon Bedrock, while Mobileye runs a support agent on Amazon Bedrock AgentCore. Both cases point to the same architecture lesson: production agents need less magic and more defined handoff points.

What changed

LendingTree describes a mortgage guidance assistant that was not built as a single chatbot. A supervisor agent coordinates two specialized workers for education and matching. The solution uses LangGraph, the Model Context Protocol (MCP), Amazon Bedrock Guardrails, Knowledge Bases, and internal APIs. Inputs and outputs pass through content filtering, PII redaction, and an additional safety classifier. According to AWS, the system has handled real conversations since late 2025; through Q1 2026 it processed around 1,960 conversations and 12,100 messages. More than 97 percent of conversations were completed without human escalation.

Mobileye shows the same point in a different environment. Its AI Support Agent answers internal status requests for the data collection pipeline. Previously, many cases required manual steps across several backends; 66 percent of support volume consisted of routine inquiries. In production, AWS reports a 98 percent success rate, response times reduced from hours to around one minute, and more than 100 tickets processed per month.

Why it matters

The relevant shift is not “more autonomy.” It is handoff architecture: Who decides which worker is responsible? Which data may a tool access? Where is state stored? Which decision can be traced later? MCP becomes the technical contract layer between agent and tool, but it is not governance by itself. Governance only emerges through routing rules, guardrails, secrets handling, logging, rollback, and human escalation.

DACH perspective

For banks, insurers, industrial companies, and public sector organizations in the DACH region, this is a useful test. Agents should not start as one large universal process, but as a chain of controlled handoffs: supervisor, domain worker, tool contract, data class, audit trail, and stop rule.

The pragmatic starting point: draw the handoffs for one use case before writing the prompts. If every handoff can be explained, the agent can become operational.