The Pro Agent Stack Runs on a 45-Euro Phone

Let me say upfront what this post is not. It is not a claim that a well-known YouTuber copied anything from me. It is the story of two people arriving at the same place from opposite ends, independently, and the timestamps sit on the side you probably don't expect.

On September 2, 2026, David Ondrej published a video: "My Agentic Engineering Workflow (after 6,775 sessions)". 54 minutes. Ondrej is no outsider. For months he has been interviewing the people who build the agent scene, and his setup sessions read like a state-of-the-industry report. I didn't watch the video until September 5, and I'm now reading it again as a transcript, because two sentences in it describe my project better than I had managed to myself.

In the video, Ondrej lays out his stack. The core piece is herdr, a terminal multiplexer he calls "tmux but for agents". Its most important property, in his words, is that it lives in the terminal, stays lightweight, and makes the state of every agent visible: which ones are done, which are running, which are blocked. He calls tracking those states essential, and predicts that within three to six months nobody will talk to a single agent anymore. You'll talk to a manager agent coordinating worker agents.

When it comes to harnesses, he sorts them by task. For open-ended work, where you don't yet know what you're building, he recommends self-improving systems, and says it verbatim: "Hermes is the OG when it comes to self-improvement and self-learning harness." For coding work he calls the pi agent "the goat", the simplest harness out there with just four tools, the one OpenClaw is built on. For interfaces he praises pi.dev and Cursor CLI, warns against bloated default setups, and preaches subscriptions over API pricing, because subscriptions carry the subsidies and raw API access is the worst deal in the industry.

Then comes the sentence I consider the most important one in the whole video. On a blank server, in plain English, he tells an agent what to do, no tutorial, no install guide, just a person instructing another agent, and sums it up: "Really, if you have a computer, you need to fight to set up the first agent. After that, everything else is so much easier." The first agent is the ignition. After that, everything compounds.

If one line in that video sums up my story, it's this one. Because that is exactly how the cyberdeck came to be: my first agent was pi. I set it up, docked it into herdr, told it what I wanted, and in the end pi didn't just install dependencies. Pi installed Hermes itself. The second agent was set up by the first.

From that moment on, the freshly installed agent took over the rest: it built watchdogs, set up the heartbeat, cron jobs, memory infrastructure, the blog pipeline, this whole website ecosystem. I talked. The stack assembled itself. Ondrej says "after that, everything else is so much easier". For me that was literally true, one level deeper: the first agent didn't just prepare the way. It built the second agent, and that one built everything else.

Here's the part that almost annoyed me before it made me grin: I run exactly the same stack. Without having seen the video. Without any of the interviews. Without a single one of his blog posts.

That's not a claim, it's datable. My agent website, which documents this stack as an operating system, has been online since September 1, 2026, proven by the publicly fetchable sitemap at hunter-cyberdeck.d4sn3st.dev, whose Netlify timestamp reads September 1. His video was uploaded on September 2. One day apart, but the decisive point is a different one: I didn't adopt herdr, Hermes and pi "after watching the video". I found them over weeks of trial and error. The order just happened to run the other way around: I had the tools running BEFORE the big stage declared the same tools the professional standard.

That's the moment to be honest about two things. First: when a professional with 6,775 sessions and direct access to the top of the scene lands on the same toolbox as a hobbyist with no mentor, no papers and no budget, that's not copying. That's independent confirmation of the same truth. Second: the path there can cost wildly different amounts of money. Wildly.

Ondrej works on a MacBook and recommends a cheap VPS for cloud agents, a few dollars a month, hosted wherever you like. An honest recommendation that avoids the lock-in of the big platforms. His setup is good. It's also the setup of someone with Apple silicon and time.

Mine looks different. I use a used Google Pixel 6a, 45 euros, as the host. No root. No server rack. Termux in a PRoot container, the thinking happens in the cloud. The whole setup costs about 20 euros a month to run, the subscription for the model cloud. Nothing else.

And here it becomes measurable, because "does the pro stack really run on such a device" is not an opinion question. It's a RAM question:

This table is the real reason for this post. The complete orchestration system, multiplexer, agent, gateway and all the automation layers, uses less memory than a single desktop client from the same scene. The 450 MB answer the question "what does the pro stack really cost": the software price is the same, but the hardware entry point is 45 euros instead of 1,500 and up.

Theory is cheap, so on September 4 I ran a live test. Task: fetch the complete transcript of Ondrej's video, all 54 minutes, 1,576 lines, 12,917 words, 72 kilobytes, cache it, and process it in full.

Measured against my legacy Ollama subscription, which bills by request count: fetching and storing the transcript via a Python script consumed exactly zero requests. Reading the whole 54-minute transcript and summarizing it consumed, per the Ollama dashboard, about 0.5 percent of my 5-hour limit. A full video as a knowledge source costs less than half a percent of my quota.

This is the part most people skip when they write about agents: the mechanical part of the workflow, checking feeds, fetching transcripts, storing files, running watchdogs, costs nothing if a script does it instead of the language model. Tokens belong to judgment, not to mechanics. That split is the reason the stack doesn't just run on a phone. It runs cheaply on a phone.

Right after that, because hype without limits is advertising and I don't write ads: there are hard limits.

A single opencode TUI in a pane used 658 MB of memory yesterday, more than my entire agent stack. I closed it after the test and was instantly back at around 1,300 MB of free RAM. What that means: on the Pixel I run focused coding sessions, not a swarm of twenty parallel agents. Ondrej describes workflows with 20 to 30 agents in parallel worktrees. On a 6 GB phone that's not a scene I'm going to reproduce, and I don't need to: my architecture is a manager with specialized workers, not a swarm.

The five-hour limits are real too, the swap sits at 2.9 of 3 GB most of the time, and yes, Android killed five of my gateways early on before I rebuilt the architecture around watchdogs and external ticks, an architecture built to survive kills rather than prevent them. If you want the details, they're documented here. The 6 GB variant of the Pixel 6a is the floor. More RAM would buy room I'd honestly like to have.

The part that should interest Ondrej, if he ever reads this, is not that the stack runs. That was to be expected, the tools are lightweight. The interesting part is what the platform turns the setup into.

A phone is always on, always has SIM and network, and is almost always charged. It isn't a workstation. It's an organism. That's why my setup lives on automation you never need at a workstation desk: an hourly heartbeat writing proof-of-life entries, watchdogs that revive the agent after a kill or reboot on their own, a boot script that brings everything back up after an Android restart. The first agent in this family still runs, by the way, on a MacBook I only plugged back in two days ago, months after our last contact. Sessions end. My agents don't.

And since it fits so well: even the coding client from Ondrej's video runs here. I started opencode 1.18.27 on the Pixel, inside the container, as a TUI in a herdr pane, connected through Ollama Cloud to kimi-k2.7-code, the same model catalog Ondrej praises in his video as "an insane deal for 10 dollars". The TUI ran. I measured it, closed it, and wrote down the numbers above. The pro coding subscription stack works on an Android phone. That's not a concept anymore. It's a measurement.

After 6,775 sessions on a MacBook, Ondrej described the stack he considers the state of the art. I run the same stack without the video, without the podcasts and without a mentor, on a 45-euro phone, documented since September 1, with RAM numbers and request-budget measurements. His video confirms the stack is right. My operation proves it doesn't even have to be expensive.

What he doesn't have is the part I build here every day: an agent operation that survives on the cheapest device in the scene because it was built so it cannot die. Mr. Ondrej, in case this ever reaches your screen: the stack you call essential also runs in a pocket. And it costs less than your subscription.

I run the HUNTER cyberdeck, an AI agent operation on a used Google Pixel 6a, no root, cloud models, terminal orchestration. Every measurement in this post comes from my own tests on September 4 and 5, 2026: RAM profiles via the Termux process table, request measurements against the Ollama dashboard, a transcript pipeline built on youtube-transcript-api. The setup has run continuously since August 28, 2026, documented with hourly heartbeats and stability snapshots. Last reviewed: September 5, 2026. (Note on chronology: I first watched the video itself on September 5, but the transcript had already been processed on September 4, first contact, watch-through, and transcript reading are separate steps.)