Post

Decagon Duet uses customer transcripts to rewrite its own operating rules without human intervention.

Most support AI relies on a human reading logs to find errors. The manager then manually updates a prompt and hopes the fix works. This creates an iteration gap where the AI stays stagnant while customers struggle.

What it does: It analyzes written records of chats to identify exactly where a conversation failed. The mechanism: It generates an Agent Operating Procedure (AOP), which is a structured set of logic rules for a specific task. Why it's interesting: Instead of a vague prompt, an AOP acts like a digital employee handbook the AI writes for itself. The safety check: The system versions the logic, meaning it saves a history of changes to prevent new errors. The result: The agent can identify a recurring payment diagnosis gap and build the fix autonomously.

This moves AI from a black box to a programmable architecture. A black box is a system where you see the input and output but cannot see how the decision was made. Programmable architecture means the agent's behavior is versioned, predictable, and measurable.

The verification process follows five concrete steps. First, it analyzes the transcript to find the gap. Second, it proposes a new AOP to solve that specific gap. Third, it creates a versioned copy of the logic. Fourth, it runs tests to ensure the change doesn't break other workflows. Fifth, it deploys the update to the live environment.

This allows a company to maintain a high-touch concierge feel even as ticket volume scales. The AI effectively manages its own quality assurance by closing the loop between a mistake and a solution.

Check out the full Field Guide PDF for the detailed breakdown.

#LLMOps #CustomerExperience #AgenticAI #CXAutomation