Roundtable recap, Oct 16: an AI assistant on a Raspberry Pi, and how much the harness matters

Most people who want an AI assistant for personal and work tasks end up paying a cloud service to hold their data. This week one of our hosts showed the alternative: an assistant that runs on a 4GB Raspberry Pi, with the data kept at home. The discussion afterward moved to a bigger question, whether the coding harness around a model matters as much as the model.

Demo: a personal agent on a 4GB Raspberry Pi

The goal was help with personal and work tasks without uploading everything to a service that costs money to keep. The first attempt used PicoClaw, a very small agent built for low-power hardware, but every tool had to be built by hand. Hermes Agent from Nous Research worked out of the box.

Setup is a one-line install. Our host recommends skipping the hosted signup wizard and configuring it yourself. The default model is Haiku 5.5 on a Claude subscription, with a few dollars of OpenRouter credit as a fallback for jobs Claude doesn't handle, such as image generation. Configuration lives in files, so the whole setup moves with you, and one update command pulls new versions without touching your skills.

The agent is reached through Telegram on a phone and through Discord, where separate channels keep topics apart and permissions stop it from seeing outside the channel it's in. It has about 67 skills and two MCP servers connected to internal tools. Browser tasks run through Stagehand, with a VNC server so a person can step in for the final click on a booking or purchase.

It hasn't been given permission to spend money or send anything on its owner's behalf. It does look after itself. When its cloud voice service started refusing requests, it fell back to a local text-to-speech engine, found a fixed library release, retested it, and asked before switching back. Memory was the one hard limit. When the Pi hit 99% memory, the agent moved stored information into skills and brought usage down to 28%.

Questions from the group

Sensitive data came up first. The data lives locally, but anything the agent works on still goes to the model provider for processing. The Pi is a test of whether the assistant earns its keep. If it does, the next step is a machine that can run a capable small open model locally.

On cost, a Claude subscription used through a long-lived token means everyday use isn't billed per token. Only the OpenRouter fallback is. One self-portrait generated with a Gemini image model cost about 4 cents.

For context in a long-running Discord channel, each new message opens a thread and a fresh session that starts with the message plus the agent's notes and skills. It fetches earlier messages with a tool call when it needs them.

Haiku handled the setup and the technical work. In our internal evals, Haiku 5.5 has passed tasks that Haiku 4.5 failed. For coding, our host still reaches for Opus.

How much does the harness matter?

One attendee uses T3 Code, an open-source app for running coding agents across worktrees that supports Claude Code and Codex in one place. For one-off tasks they see little difference between harnesses, even minimal ones. The harness matters more when ten things run at once, where reliable tools, status and retries count. Their results stay consistent partly because their MCP connections sit behind a single entry point that goes with them to every harness. They also like open-source harnesses because you can patch or fork one when a bug blocks you.

Our own evals, which run the same tasks through Pi and Claude Code, mostly overlapped. The differences showed up on the more complex tasks.

The same attendee pointed to recent reports that idle Claude Code subagents can be replaced with fresh ones that miss the prompt cache and use up limits quickly on long-running orchestration. Background processes or separate sessions are the workaround for now.

One agent at the front door

A pattern is spreading where you talk to one top-level agent that manages many parallel threads. As models get cheaper, you can point a group of agents at a whole project and only talk to the manager. It connects to graph-style setups where agents own parts of a system and hand work to each other.

When skills pile up

One attendee deleted most of their skills when they moved to newer models and kept a small set of playbooks that match kinds of work to models. Conflicting or stale skills build up fast across agents. We plan to pick this up in a later session.

The Econify.AI Roundtable meets Fridays at 11am ET / 4pm UK. Details and registration are at econify.com/ai-community.