Blog

AI agents, RAG, and integrations — notes from production.

Streaming Tokens While Running a Tool Call: Keeping Support Chat Live

A streaming AI support agent sends tokens as they generate and pauses mid-stream to run a tool - so the chat forms immediately instead of freezing on a lookup.

One Schema Per Tool: Stopping a Malformed Refund Call Before It Runs

AI agent tool schema validation: one typed schema per tool, shared by the model spec and the executing endpoint, so a malformed refund call can't be emitted.

Read vs Write Tools: Designing an AI Agent's Permissions Around Money

The AI agent read vs write tools split drives the safety design: write tools get tighter thresholds, guardrails, idempotency, and an audit record.

Containment Rate for a Voice Agent — and Why It's Easy to Game

Voice AI containment rate is the headline metric and the easiest to game - by dead-ending hard calls, by confident wrong answers, or by gaming the denominator.

How a Voice Agent Authenticates a Caller Before Taking Action

Voice AI caller authentication uses two zones: general info before verification, account-specific actions after, with the bar set by what the action risks.

Measuring Speech Recognition Accuracy by Accent and Language

Voice AI accent accuracy: an aggregate recognition score hides underserved segments. Measure word error rate and field capture per accent group, not blended.

After-Hours and Overflow: The Lowest-Risk Way to Start With Voice AI

Voice AI after hours is the lowest-risk first deployment: the alternative is a missed call, no headcount to displace, and the calls are simpler.

Function Calling on a Live Call: Keeping Lookups Under the Conversation Budget

Voice AI function calling latency is bounded by conversational turn-taking, not the database: the tools must be narrow indexed point lookups.

The Warm Handoff: What Data a Voice Agent Passes to a Human

A voice AI warm handoff transfers the live call with the transcript and verified data already on the human's screen, so the caller never repeats themselves.

Barge-In: How a Voice Agent Handles Being Interrupted Mid-Sentence

Voice AI barge-in lets a caller talk over the agent and be understood: it needs voice activity detection, endpointing, and cancelling in-flight output.