Agentcy follows a small-core architecture. The default path should be cheap, local, repeatable, and easy to inspect. Search, forecasting, publishing, live providers, and optimization loops are extensions, not the baseline.
- Karpathy's LLM Wiki pattern separates raw sources, derived wiki, and schema. Agentcy maps this to human-owned brand kits, generated artifacts, and protocol/agent instructions.
- Karpathy's autoresearch pattern keeps a fixed evaluation harness, one clear mutable surface, one metric, and a keep-or-discard loop. Agentcy maps this to deterministic artifact generation first, then optional measured optimization of one mode at a time.
Sources:
- https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
- https://github.com/karpathy/autoresearch
| Layer | Agentcy Form | Rule |
|---|---|---|
| Raw sources | BRAND.md, DESIGN.md, assets, source docs |
Human-owned. Do not mutate during generation. |
| Compiled artifacts | brief.v1, Studio inspect output, rendered cards, reports |
Agent-owned. Regenerate from sources. |
| Schema | AGENTS.md, skills, protocol schemas, docs |
Defines how agents operate and how artifacts are shaped. |
| Experiment surface | One prompt/template/mode at a time | Optimize only with a fixed eval and a clear keep/discard rule. |
agentcy pipeline run is the essential path:
brief input + brand kit -> brief.v1 -> Studio draft/render -> inspectable artifacts
It does not run Voice, Forecast, Measure, provider-backed image generation, or publish by default.
Use opt-ins for heavier work:
--personaand--persona-evalfor Voice--with-forecast --files ...for Forecast--publishfor review/publishpipeline updateandpipeline studyfor Measure after real outcomes exist
If a capability does not improve the repeated quality of artifacts from the same brand/design/content foundation, it does not belong in the core path. It can exist as an extension only when it declares its inputs, output artifact, provider cost, and verification command.