Best Practices in AI Work | Anssi Nurminen, Lasse Mikkonen | Neuvottelija 408

Summary

Three AI workers, three architectures. Anssi Nurminen and Lasse Mikkonen belong to the same fourteen-person AI working group as the host, alongside fourteen AI agents that the members can address across one another. Nurminen — speaking as a private individual, although he works in marketing at Kespro — runs Anneli, a cloud agent that turns WhatsApp photos, videos or plain text into conceptualised videos through MCP connections, mainly to Scenario. Mikkonen, a long-time DevOps practitioner who now builds agentic systems for companies, runs Ossi on the lightest rentable AWS Linux instance as pure orchestration, steered over WhatsApp, and Dude on a home machine with a 16 GB Nvidia card. The host’s Samantha runs on a Mac Mini M4.

Three terms the rest depends on. MCP is, in the host’s analogy, the USB of the database world. A skill is a way of working taught to an agent rather than a program. And the harness is the part of the system that keeps the agent on track and checks the result — Mikkonen equates it with automated testing in DevOps. What is new is that the supervision is itself built as skills.

What the agents actually do. Mikkonen’s examples are deliberately mundane: reconcile a credit-card bill by reading email, fetching the receipts and matching the PDF from the accounting system line by line — with the VAT step still missing, which he says openly. Dude’s largest skill is a case memory of projects, investments and meetings that removes note-taking. Nurminen’s distinction is the key claim: almost anyone can do one-shots with AI, but repeatable, conceptualised processes are markedly harder.

Memory is the real subject. The host describes a layered model — a lossless context expander, a wiki-like long-term memory in Google’s Open Knowledge Format, and a graph memory at the bottom. Mikkonen starts from the other end, with a graph memory built on the open-source Graphify and his own case template, and separate vector databases for large document sets. The most usable advice follows at once: ChatGPT, Claude and Gemini already have memory, steered in plain language.

Three disagreements, left open. Mikkonen runs inference in the cloud, Nurminen stays in the cloud for access to the newest video models, and the host buys his own hardware for independence should models be export-restricted. Nurminen says he has no worry even if everything leaked, because his use is pure hobby and learning; the other two hold client material. And Nurminen argues against waiting for finished tools — the jump is brutal and you miss the journey — while Mikkonen makes the same point about code: to an outsider, vibe coding and agentic engineering look alike, and the difference is security and scalability.

Meta Ray-Ban glasses are demonstrated and fail usefully: face recognition does not identify the host, biometric locking is on the device, and Finnish transcription is not supported at all. Continuous recording as memory augmentation is raised as a use case; the privacy of bystanders is raised and not resolved.

A note on the source

The summary rests on the uploader’s own Finnish caption track, which is far cleaner than machine captions. Product and model names are given as spoken in the episode; this page does not verify their features.

Watch

The recording lives on the Neuvottelija channel: Tekoälytyön huippukäytännöt | Anssi Nurminen Lasse Mikkonen | Neuvottelija 408. A Finnish edition of this episode is published at www.neuvottelija.fi.

In depth

The Neuvottelija AI editions carry a long-form write-up of this episode: English · suomeksi.