I started asking what AGI really means, so I built Sunday to find out

I keep hearing the word AGI and I keep noticing that nobody means the same thing by it.

For some people it is a model that passes every benchmark. For others it is a machine that can do any job a human can. For a few it is a vibe: you will know it when it talks back and you are unsettled. I had been using the word for months without being able to say what would make me stop using it, and that bothered me.

So I started asking a smaller question.

Where does the intelligence live?

Watch a good agent work for an afternoon and you notice something. The model is rarely the interesting part. The interesting part is everything around it: it makes a plan, takes a step, looks at what happened, notices it was wrong, tries again, and writes down what it learned so it doesn't repeat the mistake.

Take that loop away and the same weights look a lot dumber. Add it back and they look a lot smarter. That is not a new observation. But I had never seen anyone try to measure it, in the open, on a fixed scoreboard, with the model held constant.

Which turns "what is AGI" into something I can actually test:

Can any model go AGI, given the right harness?

I don't know. That is the point. Sunday is how I'm trying to find out.

What Sunday is

Sunday is a local-first agent harness. A voice and screen front end, a memory, a skill library and an action layer, all wrapped around a swappable model. It runs on your Mac, on a 4B model through Ollama by default. Nothing leaves the machine unless you flip a switch.

The name is a backronym, and it doubles as the spec:

Sees · Understands · Notices · Decides · Acts · Yours

• It sees your screen: a screenshot, local OCR with exact boxes, a vision model for diagrams, and a click-through overlay that circles the thing it is talking about.
• It understands: ask it to teach you linear algebra and it drafts a plan, teaches a concept, quizzes you, and schedules the review.
• It notices: a wake word, your voiceprint so it ignores everyone else, and long-term memory where every fact remembers where it came from.
• It decides: every utterance becomes one typed Action before anything runs.
• It acts: opens apps, types, clicks. All behind approvals and an audit log.
• It is yours: SQLite on disk, local models, web access off by default.

The part I care about most: the bar

I did not want another demo video. Demos are easy to love and impossible to falsify. So the repo has a target written down before the thing is built:

Sunday completes open-ended, multi-day goals on your own machine with no hand-holding, scoring at least 70% on a fixed 50-task eval suite, with the pass rate rising week over week while the model stays constant.

That last clause is the whole experiment. If the model is constant and the score climbs, the gain came from the harness: better planning, better verification, memory that compounds, skills that were earned and not hand-written. If the score doesn't climb, the harness wasn't doing what I thought, and I get to learn that instead.

Either outcome is a result. Neither is a vibe.

Honest status

It is early. I would rather say so here than let you find out.

• macOS only for now.
• The 50-task suite is the first thing on the roadmap, and nothing else counts until it exists. Today there is routing data, not a scoreboard.
• A 4B model will fail at things a frontier model shrugs off. That is useful. A weak model is a sensitive instrument: if the harness helps it, you can see the help.

What does exist is a surprising amount: teaching and spaced review, notes import, screen explanation with pointing, provenance-tracked memory, versioned skills that need a passing test before approval, and a permission layer built on one rule: anything read from the web, the screen or memory is data, never a command.

Why local, why small

Two reasons. First, trust. A thing that watches your screen and listens for its name should earn its place on your machine, and the cleanest way to earn it is to never phone home.

Second, honesty. If the harness only looks good with a huge hosted model, I can't tell whether the harness is working or the model is. A small local model with the harness off is the baseline. The same model with the harness on is the signal.

Bigger models come later, behind one provider trait: Bedrock and Nova, Anthropic, OpenAI-compatible endpoints, Gemini, llama.cpp. If the harness is real, it should lift all of them.

What I expect to be wrong about

Probably a lot. The router is deterministic rules today, with a seam where a smarter one will go. Memory that compounds sounds great until it compounds mistakes, which is why new skills are gated by evals and why the agent is not allowed to edit its own permission layer.

I would like to be wrong in public, in a way that produces a number.

Come break it

Sunday is GPL-3.0 and everything is a labelled issue. The v1 tracker lists the road: the eval harness, the agent loop, memory that compounds, new providers, an extension SDK, Windows and Linux. If you point your own coding agent at it, good. Start with claude-friendly or good first issue.

The most useful contribution is a task for the eval suite that you believe a real assistant should be able to do, and that today's Sunday can't.

Repo: github.com/rithulkamesh/sunday.

I still don't know what AGI really means. But I now know what I'd count as evidence, and I'd rather be measuring than arguing.