Blog
Notes from the Vantaso team
Field notes from the Vantaso team on shipping AI-driven software — engineering practices, LLM architecture, and the products we build for clients and ourselves.
Agent Sandbox Escape: The Boundary Was Never the Box
Two labs published agent sandbox escape postmortems in 2026. No container was broken into. What set the blast radius was access, grading and monitoring.
Verbatim Speech-to-Text: Every Engine Corrected Our Learners' Grammar
We gave one learner recording to ten speech-to-text engines and told them not to fix anything. All ten repaired grammar. Here is what survived.
Agent Context Compaction Quietly Deletes Your Guardrails
Agent context compaction drops the standing rules you load into context: measured violation goes 0% to 30%. Where to pin a rule, and where to enforce it.
Stateless MCP: Your Server Just Inherited the State
Stateless MCP dropped the handshake and Mcp-Session-Id in the 2026-07-28 spec. What it hands back: explicit state handles, expiry, per-request auth.
LLM Regression Testing: What Breaks When You Swap Models
LLM regression testing for model swaps: the silent changes a correctness eval misses, what to measure, and why a falling score can mean a better model.
LLM Provider Routing: What a Gateway Won't Fix
LLM provider routing makes a model swap a config change. But a gateway won't fix cache locality, tokenizer drift, or silent quality regression.
Web Scraping With LLMs: Markdown First, Then Structure
Web scraping with LLMs: scrape each page to clean markdown, then extract structured data against a JSON schema and validate it. With real token math.