Lab notes

What we ship, with the numbers.

September 9, 2026 We put an agent on a club's LINE account. Escalations go to two managers' groups and come back as replies A multi-location club handed us a sixteen-module spec for a LINE concierge. It went live in a morning as a scene, two subscriptions and one rule in the webhook. Here is the build, what was silently broken before it, and how renewals run off the booking system.
September 9, 2026 An agent reschedules a meeting across Google Calendar and Outlook A partner proposed a new time. The agent found it, checked both calendars, and answered. Two things got in the way, a field Google's API does not have and a calendar set to Asia/Qatar.
August 27, 2026 Why we stopped paying for DocuSign: PAdES digital signatures in 40 lines of Python Electronic signatures don't need a $40/seat SaaS vendor. The PDF standard has included open cryptographic digital signatures (PAdES) for decades. Here is how to build an agentic signing pipeline.
August 27, 2026 We automated our Fortune 50 vendor onboarding form with a single agent Every enterprise vendor onboarding has "that form" — a multi-page PDF questionnaire with custom security ID formulas, checkboxes, and baseline signatures. We had an agent handle it in 60 seconds.
August 14, 2026 The same open-weight model costs up to 12x more depending on which host answers. The expensive hosts are not serving better weights. Gemma 4 31B is served by 19 hosts on OpenRouter at prices from $0.08 to $0.99 per million input tokens, and the cheapest one serves higher precision and double the context of the most expensive. The router picks per call, so most people are paying a number they never chose.
August 7, 2026 AI security agents patch nine of ten known vulnerabilities. They find one in ten. Cybersecurity's main agent benchmark went from 5% to 100% in two years. The one that replaced it splits the work into detect, exploit and patch, and the gap between those columns says more than any ranking.
August 7, 2026 The top document parsing models score 96 out of 100. No benchmark checks whether they read the invoice total correctly. The leading document parsing models sit within a point of each other above 95. We went looking for one that could read a scanned purchase order in Thai, found the leaderboards couldn't answer, and working out why was more useful than the rankings.
July 27, 2026 What we learned putting AI glasses on a robot bench We wanted to know what AI glasses are good for when the thing you're looking at is a machine and the result has to land in an operational system, not a chat window. So we built the loop on our own arm and wrote down what the bench taught us.
July 18, 2026 The best Thai speech-to-text models drop every English word We benchmarked five speech-to-text systems on Thai — commercial APIs and open models, scored the same way. The ones that nail Thai characters throw away the English, and one common default fails outright.
June 2, 2026 Matching NVIDIA's flagship PII detector with a model 3× smaller A 50M-parameter model ties NVIDIA's gliner-PII on the same benchmark, trained in 39 minutes for ~$8 — a reproducible demonstration that a task-specific model can be far smaller than the general one.
May 25, 2026 Every agent needs a wallet Agents need identity and credentials. We built kern — an open-source TypeScript SDK for agent secrets, built on age encryption.
May 25, 2026 MCP Apps are here. Your tools can ship UI into Claude and ChatGPT. MCP tools can now return interactive UI that renders inside AI clients. Here's how it works, what it means, and what's missing.
May 25, 2026 Elicitation: MCP's standard for human-in-the-loop MCP servers can now ask users for input mid-execution. Here's the protocol and why it matters for agent trust.
May 25, 2026 We screenshot our Windows app from CI. Here's how. How we built an e2e smoke test that installs our Tauri app on Windows, logs in with real credentials, and captures screenshots — all in GitHub Actions.
May 25, 2026 A2UI vs MCP Apps: two standards for agent UI, and why you'll need both Google's A2UI ships declarative JSON. MCP Apps ships full HTML. Here's when to use which.