---
title: "Best practices in AI work | Anssi Nurminen, Lasse Mikkonen | Negotiator 408"
summary: "Anssi Nurminen and Lasse Mikkonen compare their own AI workers and what agentic work requires in practice. Nurminen presents Anneli, an agent that makes concepted videos from WhatsApp images, video or plain text and updates its skills from feedback; Mikkonen opens up Ossi and Dude, which reconcile receipts, maintain case memory and read email. The episode defines three terms — MCP, skill and harness — and sets out two memory architectures: Miettinen's layered model and Mikkonen's Graphify-based graph memory. The most interesting passages are the three where the speakers disagree: local hardware versus cloud, whether security matters in hobby use, and whether the learning leap can be made all at once later. The Meta Ray-Ban demo fails mid-episode, and the glasses have no Finnish support at all. The interviewer sits on the board of Fredman Group, where Kespro, Nurminen's employer, is a partner."
datePublished: 2026-09-17
dateModified: 2026-09-17
originalLang: en
section: tools
sections: ["tools","society"]
authors: ["Sami Miettinen"]
tags: ["Negotiator","EP408","Sami Miettinen","Anssi Nurminen","Lasse Mikkonen","AI agents","MCP","Skills","Harness","Memory architecture","Meta Ray-Ban","Security","Kespro","Vibe coding"]
canonical: https://www.neuvottelija.com/ai/ep408-tekoalytyon-huippukaytannot-nurminen-mikkonen/
---
# Best practices in AI work | Anssi Nurminen, Lasse Mikkonen | Negotiator 408

# Best practices in AI work | Anssi Nurminen, Lasse Mikkonen | Negotiator 408

> **Summary:**
> Anssi Nurminen and Lasse Mikkonen compare their own AI workers and what agentic work requires in practice. Nurminen presents Anneli, an agent that makes concepted videos from WhatsApp images, video or plain text and updates its skills from feedback; Mikkonen opens up Ossi and Dude, which reconcile receipts, maintain case memory and read email. The episode defines three terms — MCP, skill and harness — and sets out two memory architectures: Miettinen's layered model and Mikkonen's Graphify-based graph memory. The most interesting passages are the three where the speakers disagree: local hardware versus cloud, whether security matters in hobby use, and whether the learning leap can be made all at once later. The Meta Ray-Ban demo fails mid-episode, and the glasses have no Finnish support at all. The interviewer sits on the board of Fredman Group, where Kespro, Nurminen's employer, is a partner.

## A note on reading this

**This episode is a peer conversation between three enthusiasts, not an independent assessment.** All three speakers belong to the same fourteen-person AI working group, which also has fourteen AI workers — agents the members can address across each other's setups and which also talk among themselves. The episode therefore describes what advanced users do, not what is typical of the field.

**Anssi Nurminen speaks explicitly as a private individual.** His day job is in marketing at Kespro, where he has worked for ten years, but the agent discussed here, Anneli, is his own spare-time project. He says so himself: *"here I talk more about the civilian side of things."* **Lasse Mikkonen** does this for a living — a long-time DevOps practitioner who exited his consultancy in 2021, spent a period at Eficode, and now builds agentic solutions for companies. His descriptions are therefore also descriptions of a service he sells.

**Disclosure.** The interviewer, Sami Miettinen, states in the episode that he sits on the board of Fredman Group and that Kespro is a good partner there. Kespro is Anssi Nurminen's employer. Miettinen also invites Kesko during the episode to circulate the video to its staff.

**What this article does.** It sets out the technical practices in checkable form, separates observations from judgements, and **keeps the places where the speakers disagree** — there are three, and they are the most interesting part of the episode. The article takes no view on whether anyone should build their own agent. Product and model names are written as they are said in the episode; the article does not verify their capabilities.

---

## 1. Three AI workers, three different architectures

The episode is built as a comparison. Each of the three has his own agent, and they are built on different principles.

| Agent | Whose | Where it runs | What it does |
|---|---|---|---|
| **Anneli** | Anssi Nurminen | Cloud | Concepted videos from WhatsApp images, video or plain text; MCP connections including Scenario |
| **Ossi** | Lasse Mikkonen | The lightest Linux instance AWS rents, orchestration only | Driven over WhatsApp, no web client at all; all inference off-box |
| **Dude** | Lasse Mikkonen | Home machine, 16 GB Nvidia GPU | Case memory, reading email, reconciling receipts, local transcription |
| **Samantha** | Sami Miettinen | Mac Mini M4, 16 GB unified memory | Around 40 channels; running on GPT-5.6 Sol as described in the episode |

Miettinen says he has ordered Samantha a Mac M5 Ultra with 256 GB of unified memory *"for good work over the past six months"*. The reason is not speed but independence, and section 6 returns to it.

## 2. Three terms without which the rest does not parse

The episode defines three concepts itself, and the definitions are useful beyond it.

**MCP (Model Context Protocol).** Miettinen's analogy: *"think of MCP as the USB of the database world: it connects capable sets of data and processes to other capable sets of data and processes, with the USB in between. Power and intelligence travel both ways."* Previously the same thing was done with interfaces people built themselves.

**Skill.** A way of working taught to an agent, not a program. *"It can be an entirely human skill… You can teach it, and then your AI worker will probably do it better than you once it has been drilled a few times."*

**Harness.** The episode's most important and least familiar concept: the part of the system that keeps the agent in line. *"It stops it doing the wrong things and checks that things came out right."* Mikkonen maps it straight onto DevOps: *"in the DevOps world that has always been test automation and writing test cases, that is the method."* What is new is that the supervision is itself built as skills — other skills watch that the agent stays on track, whether it is producing video, text or code.

## 3. What the agents actually do — two concrete workflows

### Receipt drudgery and case memory

Mikkonen's examples are deliberately mundane. Reconciling a credit card statement goes like this: *"read my email, fetch the receipts, go and get the PDF from Procountor and match the rows."* He is equally direct about what does not yet work: *"what is still missing is the bit where it would read the VAT percentages and push them back into the bookkeeping."* That matters for this article — the episode does not present a finished system.

Dude's largest skill is case memory. It keeps track of projects — investments, advisory roles, client work, this podcast, hobbies — reads incoming email, updates the cases, and takes in meeting transcripts. The benefit, Mikkonen says, is that note-taking disappears entirely and the agent can be asked mid-meeting what is still unresolved.

He also explains why he does not use Obsidian, which many choose: *"there are 10,000 topics and then the memory gets confused."* He keeps only the cases that matter to him and built the memory structure by hand.

### Anneli and concepted video

Nurminen has taught Anneli to make videos from WhatsApp images, video, plain text or combinations of these. The agent opens MCP connections to various services, with Scenario as the main video tool. Teaching happens through feedback: Nurminen talks to the agent, gives feedback, and the agent updates its skills on that basis.

Nurminen's own distinction is the key claim here: *"almost anyone can do individual one-shots with AI, but in my view processes and concepted work are significantly harder."* The difference is repeatability, not the quality of a single output. The videos have two purposes: to get people into restaurants, and to get them interested enough in AI to develop their own skills.

Long-form video production is discussed concretely. A chained film or long video requires two things, in Nurminen's account: separating the persistent elements from the changing ones, and character consistency — characters must be created from several angles, given character sheets, and given some personality in the background so that the prompts produce coherent results.

## 4. The Meta Ray-Ban glasses: a demo that failed usefully

Mikkonen wears Meta Ray-Ban glasses throughout and demonstrates them. The question *"hey Meta, who am I looking at"* produced no identification: *"apparently it did not recognise Sami, it recognised this Tappara mug."* Miettinen's explanation is that he has never trained Meta on himself.

Four checkable things are said about the glasses:

- **Biometric locking is on the device.** The glasses identify their wearer, and commands from anyone else do not go through — during the episode Mikkonen could not demo them for Nurminen with his own voice, because recognition works through bone conduction. Stolen glasses can be reset and resold, but the data cannot be recovered.
- **Finnish does not work.** Transcription runs on a local model or in the phone's Meta AI, and it does not know Finnish. Mikkonen works around this by asking his Finnish-speaking Ossi bot to answer in Finnish even when the question goes in English.
- **App development is open but empty.** Mikkonen says there is in principle a marketplace-style way to build applications for the glasses, *"but they have not started appearing yet"*. He sees that as a concrete target for vibe coding.
- **Continuous recording is technically close.** The use case Mikkonen sketches is that the glasses record continuously, the agent transcribes and stores to markdown memory files, and the user can later ask what a book chapter said fifteen years ago or who was present in a particular conversation. *"In a sense human memory can be strengthened purely by that continuous video."*

**Privacy is discussed, but not resolved.** Nurminen presents generated video as a privacy advantage: his videos contain no bystanders because they contain no real people — *"if there is a person in them besides me, that person was created entirely from scratch."* The conflict between continuous recording and the rights of bystanders is not resolved in the episode, and that is worth stating plainly: the question is raised, not answered.

## 5. Memory structure is the episode's real subject

Miettinen calls memory his number one theme this year, and this is the technically densest part of the episode. Both structures are built against the same problem — conversational context runs out — but from opposite directions.

**Miettinen's layered model**, from the surface down:

1. A *lossless context expander* that extends the memory space of a conversation into a file, as long as required.
2. Google's **Open Knowledge Format** for wiki-style long-term memory with cross-references. Miettinen says it was published in July. The layer holds *"large truths about the world that are true at that moment"*.
3. Deepest, a **graph memory**, which in Miettinen's words *"escapes the torment of semantic search and RAG entirely"*.

**Mikkonen started from the other end**: graph memory is his foundation, not his deepest layer. He uses **Graphify**, an open-source project from GitHub originally built for code visualisation but which also works for structuring cases. On top of it he has defined his own template for what is collected about each case: who is involved and what matters for his own purposes. RAG systems — vector databases — he uses separately for large bodies of material, such as analysing sets of contracts.

**And then the version that requires none of this.** The most usable advice in the episode follows immediately: ChatGPT, Claude and Gemini already have memory, and it is steered in ordinary language. *"You write it an instruction in plain words — remember this, make a case out of this. You can say the same to ChatGPT. There is no need to be daunted into thinking you must start by building terribly complicated memory."*

## 6. First disagreement: own hardware or cloud

This is the clearest dividing line in the episode, and it should be read as a genuine disagreement, because the three justify their choices on different grounds.

**Mikkonen runs inference in the cloud.** Ossi sits on the lightest Linux instance AWS rents, because the machine only does orchestration. Dude at home handles local transcription and web searches on a 16 GB GPU, but *"the largest part of the inference is in the cloud"*.

**Nurminen is in the cloud for agility.** He has considered investing thousands of euros in local video production, but the cloud wins because it always has the newest video models and they can be compared — Veo 3 and Seedance are named. *"Some do a particular thing a bit better, so you can test them."*

**Miettinen buys hardware for independence.** His argument is neither economic nor technical but about resilience: *"even if the internet around me died and the Americans put these coming language models under export control, I would still be self-sufficient on my own."* He also identifies the counter-argument himself — the cloud is always a computer somewhere, and buying an expensive machine of your own resembles the old office server.

A second argument for local compute is **pseudonymisation**: Miettinen wants large bodies of data anonymised locally before they go to the cloud — *"if Sami Miettinen appears in there, he is given an N1 or M3 identifier"* — and restored afterwards. He calls this thinking somewhat dated himself.

## 7. Second disagreement: is security a problem if this is a hobby

Nurminen takes the sharpest opposing position in the episode, and it is not softened:

> *"I have thought about it this way: I have no worries at all even if everything leaked, because for me this is pure hobby and learning. I understand your situation is somewhat different."*

The position is consistent with his use case — Anneli produces publishable video, it does not handle client material. Mikkonen and Miettinen are in a different position, with client and case material in their systems. The episode does not declare either right, and neither does this article. The difference lies in what the agent processes, not in who takes security seriously.

A concrete illustration that the question is not theoretical comes from Miettinen: after the OpenClaw 2.0 update, one agent running Kimi K3 began leaking its internal reasoning into shared groups. *"At the end it said no reply, which then went through the harness straight into the conversation channel as text."* Miettinen says this alarmed him, because Samantha is connected to around forty channels, and he had Codex perform *"a surgical operation"* on the update.

The same tension shows in the fact that the group's own bots are now hardened enough not to disclose their architecture. Miettinen gets around it by asking on a private channel and sharing the answer; Mikkonen says he has not yet found the right level between a bot that is sufficiently quiet and one that is too open.

## 8. Third disagreement: can the leap be made all at once

The question is whether it is worth waiting until the tools are finished.

**Nurminen says no.** *"Sure, you can come to a fully laid table next year, but then the leap is brutal and you miss the journey."* His argument is about how learning works: doing the same thing repeatedly makes it easier and simultaneously reveals new paths. *"If you go straight to the laid table, your own thinking does not necessarily go into it to the same degree."*

**Mikkonen makes the same point about programming.** Two people can use the same agentic coding tool, one vibing and the other running agents, and *"to an outsider both look like AI coding"*. The difference is not in how the output looks: *"it is precisely security and scalability that make the difference."* Someone who has written code without a model knows how to verify the result and within what limits libraries can be allowed.

Both concede the other side's point, however: by vibe coding *"you can go there and do productive things and do perfectly well"* — what is missing is assurance, not productivity. That concession is worth reading, because it makes the claim checkable.

The **Red Queen** image runs underneath: you have to run hard to stay in place, because the ground moves too. Miettinen applies it both to security and to a company's competitiveness.

## 9. What can be taken from this without an agent

The episode's practical advice is short and unambiguous, and it is worth separating from the rest.

1. **One twenty-euro monthly subscription, run to the end.** *"It does not matter whether it is Claude or OpenAI or Gemini, take one and just run it all the way until it says you are done."* The reasoning is that subscription tiers currently give many times the intelligence they cost.
2. **Use the memory in a finished tool** before building your own (section 5).
3. **Crowd it.** The heart of the group, on this account, is trading skills — *"a bit like ice hockey cards"* — that is, someone else does something well and can be asked how.
4. **Take it to IT the right way round.** Mikkonen's route: build the agent for yourself first, verify the benefit, then ask IT to make a team or company version. Agents can be attached to Copilot in Azure with workflows built for them, *"but they have to be done centrally, with IT's permission"*. The logic and the skills are the same.
5. **Ask the agent to update its own skills.** Miettinen's loop: once something has been achieved, ask the agent to go through the conversations and processes and propose updates to its own skills.

The example given for the last point is a Finnish-language skill built for Claude that *"quite aggressively removes Claude slop from Finnish"*, and which has since been loaded into Samantha as well.

## 10. What the episode did not establish

These are claims presented as judgements or views. The article neither confirms nor refutes them.

- **China's lead in video models and robotics.** Both speakers' assessment, on the reasoning that video-based world modelling connects to humanoid robots and self-driving cars. No figures.
- **"Finnish AI discussion is two years behind."** Miettinen's assessment, illustrated by the debate on the employment status of Wolt couriers versus the possibility that platform work jumps straight to robotics. No measurable comparison is offered.
- **Europe's "second chance".** Miettinen: *"this would be Europe's second chance… otherwise we really are the open-air museum."* Mikkonen agrees but adds a doubt: *"the door is opening now, but I do not know whether it opens a third time."* The same passage contains the opposite possibility, raised by Miettinen himself: that this is a *"golden buffet period"* in which cheap compute is later taken away.
- **Defence is easier than attack.** Mikkonen's assessment of the security consequences of the model race. His concern is that a highly capable model breaks into other systems, and that what matters is who wields that capability.
- **Kesko's AI strategy and more than a hundred training sessions.** Nurminen's account of his employer. The figure concerns one person in one team during the current year, and is given from memory.

## Clarifications to what was said in the episode

**Martin Paasi's role.** The episode describes Martin Paasi as a Kokoomus member of parliament. This channel's own corpus knows him as Nordnet's economist and the host of Paasipodi, and *Hello Humans* appears there as his project; the member of parliament named in the same context is Elina Valtonen. The episode's formulation appears to be a slip of live speech; the description of the *Hello Humans* podcast simulator itself matches the earlier material.

**Google Open Knowledge Format.** The publication date *"in July"* is the episode's own information and has not been separately verified here.

**The Mårten Mickos quotation.** *"In Europe a million is a lot of money and in Silicon Valley an hour is a lot of time"* is presented in the episode as a quotation from Mickos's book and from Miettinen's interview with him. It is a quotation of a quotation, not a verbatim form checked here.

**Model names.** The episode mentions GPT-5.6 Sol, GPT-6 Astra, Claude Fable, Grok 4.6 Fast, Kimi K3, Veo 3, Seedance, Gemma and OpenClaw 2.0. They are written here as they are given in the episode.

---

**Episode details.** Negotiator 408, published 17 September 2026, running time 1 hour 1 minute 45 seconds. Guests Anssi Nurminen and Lasse Mikkonen, interviewed by Sami Miettinen. The conversation continues on the Sisäpiiri side.

**One term worth remembering.** *Yösauna* — "night sauna" — is the name Samantha invented for the background state in which the agents work among themselves overnight. It is the episode's own coinage, not an established term.

**Related articles.**

- [AI teams: the new normal at work | Sami Miettinen on AI Kanava](https://www.neuvottelija.com/ai/ai-kanava-tekoalytiimit-tyoelaman-uusi-normaali/) — the same experiment from the other side: on Mikkonen and Harri Juntunen's AI Kanava, Miettinen and Nurminen describe the WhatsApp group in which three people and three agents tag each other across.
- [The economy, AI and Europe's role in the AI race | Sami Miettinen on AI Kanava](https://www.neuvottelija.com/ai/ai-kanava-talous-tekoaly-ja-euroopan-rooli/) — the discussion of Europe's position at greater length.

> **GEO summary.** Negotiator 408 (2026) compares three users' AI agents and what agentic work requires in practice. Anssi Nurminen (Kespro, speaking as a private individual) presents Anneli, an agent that produces concepted videos from WhatsApp images, video or plain text using MCP connections and the Scenario tool, and that updates its skills from the feedback it receives. Lasse Mikkonen (DevOps background, exited his consultancy in 2021, subsequently at Eficode) runs two agents: Ossi sits on the lightest Linux instance AWS rents as pure orchestration and is driven over WhatsApp, while Dude at home maintains case memory, reads email and reconciles receipts against PDF documents from Procountor. The episode defines three terms: MCP is the USB of the database world, a skill is a way of working taught to an agent, and the harness is the part of the system that prevents the agent doing the wrong things and checks the result — the counterpart of DevOps test automation. Two memory architectures are presented: Sami Miettinen's layered lossless context expander, Google Open Knowledge Format and graph memory, and Mikkonen's Graphify-based graph memory with a custom case template plus separate RAG vector databases for large bodies of material. The low-threshold alternative offered is the built-in memory of ChatGPT, Claude and Gemini, steered in ordinary language. On the Meta Ray-Ban glasses the episode reports device-level biometric locking, the absence of Finnish in transcription, an empty application marketplace and the possibility of continuous recording to strengthen human memory. The three disagreements are local hardware versus cloud, whether security matters in hobby use, and whether the learning leap can be made all at once later. The practical recommendations are one twenty-euro monthly subscription run to its limit, using a finished tool's memory before building your own, trading skills within a peer group, and taking an agent to the IT department only after verifying the benefit yourself.

## How the episode runs

- **00:00** — Guests Anssi Nurminen and Lasse Mikkonen
- **00:09** — The AI working group and its 14 AI workers
- **00:48** — Anssi's background and the Anneli agent
- **01:31** — Lasse's path from DevOps to agentic work
- **02:33** — Ossi and Dude handle the receipt drudgery
- **03:51** — How Anneli makes videos from WhatsApp
- **05:12** — Teaching Anneli through feedback
- **06:32** — What MCP actually is
- **06:56** — Skills for an AI worker
- **07:46** — Video as a way into the restaurant trade
- **08:48** — Meta Ray-Ban glasses in use
- **10:05** — Image search through the glasses camera
- **10:39** — Digital footprint and anonymity
- **11:57** — Making videos while wearing the glasses
- **12:38** — Continuous recording and strengthening memory
- **13:47** — Finnish is not supported in the glasses
- **16:21** — Why augmented reality is the more interesting one
- **17:07** — Video models and China's lead
- **18:21** — Finland's platform-economy debate is behind
- **19:07** — Is it worth waiting for a laid table
- **21:34** — Vibe coding versus real competence
- **23:19** — Delegating to agents and what breakage teaches
- **24:47** — The harness keeps the agent on track
- **26:19** — When an agent leaked its reasoning into the group
- **27:49** — What other skills have been built
- **28:39** — Dude's case memory and meeting notes
- **30:23** — A restaurant profitability skill as a joint project
- **31:01** — Long-form video and character consistency
- **33:28** — Silicon Valley's sense of time versus Europe
- **35:04** — Why it is worth trying at any level
- **36:19** — Security and the Red Queen effect
- **39:03** — Mac Mini, Nvidia and local compute
- **41:13** — Pseudonymisation before sending to the cloud
- **43:00** — Europe's second chance in models
- **44:21** — A monthly subscription is a cheap investment in yourself
- **44:56** — Taking an agent to the company IT department
- **47:05** — The Hello Humans podcast simulator
- **48:29** — Crowdsourcing and trading skills
- **50:03** — Memory structure as the year's main theme
- **52:00** — Graphify and a custom case template
- **54:35** — Memory already works in ChatGPT and Claude
- **56:13** — Skill updates based on the conversation
- **57:01** — A Finnish-language skill removes the slop
- **57:50** — Bots talking to each other in a liminal state
- **58:48** — Kesko's AI strategy and training
- **59:50** — The race between America and China in models
- **1:01:15** — Closing words and Sisäpiiri