sunday
Can any model go AGI, given the right harness?
A local-first voice and screen agent harness: plan, act, verify, remember, learn. An open experiment, scored by evals.
Demo
The question
A lot of what looks like intelligence in an agent comes from the loop around the model: planning, verifying, remembering, learning from failure. Not only from the weights.
If that is true, a small local model with a good harness should keep getting more capable, and a bigger model should get further still.
The bar
Sunday does not claim AGI. It claims a measurable target:
Complete open-ended, multi-day goals on your own machine with no hand-holding, scoring ≥ 70% on a fixed 50-task eval suite, with the pass rate rising week over week while the model stays constant.
If the number moves, the harness is doing real work. If it doesn't, that is a result too. Either way it is a number, not a vibe.
S.U.N.D.A.Y.
Sees
Screenshots with local OCR and exact boxes, a vision model for diagrams, and a click-through overlay that points at things.
Understands
Teaches any subject with plans, quizzes and spaced review. Imports your notes and cites them.
Notices
Wake word, voice activation, your voiceprint, and long-term memory where every fact has provenance.
Decides
Every utterance becomes one typed Action, checked against what is permitted. A deterministic router with a seam for a smarter one.
Acts
Opens apps, pages and files, types, clicks with a preview. All behind approvals and an audit log.
Yours
Local models via Ollama, SQLite on disk, web access off unless you switch it on.
The loop
Where it is today
Early software. macOS only, and it runs on a 4B local model by default. Next on the roadmap:
- an eval harness and the 50-task suite. Nothing else counts until it exists
- a persistent agent loop: plan, act, observe, verify, replan
- memory that compounds: successful runs become skills, gated by evals
- any model: Bedrock / Nova, Anthropic, OpenAI-compatible, Gemini, llama.cpp
- MCP tools, a plugin SDK, and Windows and Linux
Run it
ollama pull qwen3:4b-instruct && ollama pull nomic-embed-text-v2-moe
bun install && bun run tauri devWhy I Built This
I got tired of arguing about what AGI means and wanted to build something that could be wrong in public.
So Sunday is an experiment with a scoreboard. Read the story behind it, or pick up an issue and help.