Daily recap for May 24, 2026

Today we shipped four step-by-step tutorials covering AI agent reliability and voice scheduling, plus a vendor-agnostic RAG toolkit, and 13 new building blocks for conversation state management, output repair, and voice pipelines.

New tutorials

Mistral AI agent reliability suite for SMB customer support

When your Mistral-powered customer support agent can't reach a downstream API—like order lookup or the return portal—the entire agent fails, and customers get nothing. This tutorial shows you how to wrap every tool call in a circuit breaker and add idempotent retries, so one outage doesn't cascade. You'll also set up health checks and incident runbooks that kick off automatically when something breaks, keeping your support agent running regardless of which downstream service is having problems.

Read the tutorial → | Download the code (zip)

Under the hood: Mistral, reliability-ops stack with Hono, 63 tests, 98.51% coverage.

Vertex AI voice agent for Twilio field service scheduling

Small field-service businesses miss after-hours calls and spend hours manually juggling technician schedules, leading to lost jobs and double-bookings. This recipe deploys a real-time voice agent that answers incoming Twilio calls, books appointments directly into Google Calendar, and routes callers to the right specialist—booking, rescheduling, or FAQs. It stays within a per-call budget and tracks everything with observability, so you're never in the dark.

Read the tutorial → | Download the code (zip)

Under the hood: Vertex AI, voice-agent archetype with Next.js and Express, 135 tests, 98.95% coverage.

Anthropic MCP server for SMB ERP tool access

SMBs want AI agents that can pull real-time data from HubSpot or QuickBooks, but hand-building secure, authenticated, and rate-limited endpoints for each system is fragile and slow. This tutorial walks you through building an MCP server that exposes your HubSpot CRM as type-safe tools Anthropic's Claude can use, with API-key auth, per-tenant tool allowlists, token-bucket rate limiting, and audit logging all built in.

Read the tutorial → | Download the code (zip)

Under the hood: Anthropic, MCP server archetype with Hono, 90 tests, 95.50% coverage.

Vertex AI reliability suite for SMB agent operations

Small-business AI agents regularly fail when downstream tools go down, LLMs return malformed output, or retries multiply, causing customer-facing disruptions without a dedicated SRE team. This tutorial layers circuit breakers, idempotency middleware, and structured output repair onto Vertex AI calls, then wires them into runbook incident workflows that escalate automatically—giving you production-grade fault tolerance that's reusable and fully tested.

Read the tutorial → | Download the code (zip)

Under the hood: Vertex AI, reliability-ops stack with Hono and Next.js, 71 tests, 98.38% coverage.

RAG pipeline toolkit for production AI

For businesses building retrieval-augmented generation pipelines, the RAG pipeline toolkit provides vendor-agnostic, modular building blocks: pluggable embeddings, vector stores, and rerankers. Everything is written in TypeScript with ESM, ready to drop into your own stack.

Browse the toolkit →

Under the hood: RAG knowledge archetype with TypeScript, ESM, pnpm workspaces.

Building blocks shipped

Session continuity

The core @reaatech/session-continuity package manages multi-turn AI conversation state, token budgets, and context compression. The family also shipped a tokenizer package for OpenAI and Anthropic models, plus storage adapters for DynamoDB, Firestore, Redis, and an in-memory option for testing. Browse the session continuity catalog →

Structured output repair

The structured output repair family published a core package that fixes malformed LLM JSON output using a pipeline of strategies to match a Zod schema, and an MCP server exposing those repair tools over the Model Context Protocol. Browse the structured output repair catalog →

Voice agent kit

The @reaatech/voice-agent-core orchestrates STT→MCP→TTS voice agent pipelines with per-stage latency enforcement and OpenTelemetry observability. Accompanying it are providers for speech-to-text (Deepgram, AWS Transcribe, Google Cloud), text-to-speech (Deepgram Aura, Polly, Google Cloud), a Twilio Media Streams handler, and an MCP client. Browse the voice agent kit catalog →

Browse the full catalog at reaatech.com/products.