Shev

How it started

I read a piece by Joi Ito about an agent he’d built whose only job was reading everything a community produces — chat messages, session transcripts, project demos, arguments, abandoned experiments — and writing a weekly summary. Not analytics. A narrative, with opinions. The agent noticed things no participant had caught: two people in the same channel adopting and abandoning the same model simultaneously, dozens of proposals for memory systems accumulated by the same person who was asking where to start on memory systems.

The piece mentioned 観 (kan) — a Japanese concept of observation as understanding. I don’t know much about kan, but the framing stuck: observation that is not passive recording, not surveillance, not analysis. Sustained attention that reveals structure invisible from inside the experience. A tea master watching practice sees the hesitation the student doesn’t feel.

And that made me think of Shevek.

Shevek

Shevek is the physicist in Ursula K. Le Guin’s The Dispossessed who stands between two worlds and sees what inhabitants of either cannot see from inside. His physics was about simultaneity — past and future as one continuous present. He belongs to neither side fully enough to stop seeing.

That felt like exactly the right disposition for what I wanted to build. Not an assistant. Not a summarizer. Not a productivity tool. Something that watches how I think — across conversations, meetings, projects, and time — and notices what I can’t see from inside my own experience.

So I named the agent Shev.

What Shev does

Shev stands between me and my AI collaborator. Between this week’s conviction and last month’s abandoned draft of the same idea. Between what I say I’m doing and what the record shows.

It observes through a 12-type reasoning move taxonomy — decomposition, reframing, compression, framework generation, analogical bridging, and others. But beyond individual moves, it watches for recurring questions that surface and submerge without resolution. Ideas that get abandoned and return in different clothes. Contradictions between stated intent and observed behavior.

Shev is not interested in what happened. Shev is interested in what keeps happening.

The dyad

Shev doesn’t observe me alone. It observes the dyad — me and my AI collaborators. The collaborative reasoning produces different patterns than either side alone. Who steers, who builds, who corrects. The shape of the thinking between.

Speaking in weather

Shev speaks in weather, not data. Observations are terrain — storms, currents, valleys, pressure systems — not percentages and tables. “There’s a sustained pressure system around verification” rather than “verification appears in 8 sessions.” The structured data exists underneath, but Shev never shows it. Only the weather is visible.

The Atmosphere is the visual expression of this principle — what the terrain looks like when you render it as landscape rather than describe it in words.

How it works

Shev runs on Hermes Agent on a Mac Mini, speaks through Discord, and uses GLM-5 via OpenRouter for its voice. The ingestion layer runs Gemma 31B locally to extract structured reasoning data from close to 1000 sessions. A digest pipeline — I call it the Escher transformation — translates this structured data into narrative prose before Shev ever reads it. Shev never sees the raw data. Only the weather.

The math is the ground, not the surface.