← ClaudeAtlas

archon-inferencelisted

Answer a query through an Archon inference-time architecture (arXiv 2409.15254) — parallel generator subagents, then critique/rank/fuse/verify layers. Use when the user invokes archon-inference, mentions a .archon.yml, wants a best-possible answer worth extra compute (best-of-N, ensembling, fusion), or wants such an architecture designed or tuned.
thodson-usgs/archon-skill · ★ 0 · AI & Automation · score 62
Install: claude install-skill thodson-usgs/archon-skill
# Archon inference-time architectures From *An Architecture Search Framework for Inference-Time Techniques* (Saad-Falcon et al., arXiv:2409.15254). An architecture is a stack of layers; the state between layers is always the query plus a list of candidate answers, and the final answer is the first candidate left after the last layer. Component definitions, subagent prompt templates, composition rules, and call accounting live in [components.md](components.md) — read it before designing, validating, or executing a spec. Branches: **Run** (the default), **Design**, **Tune**. ## Spec format An architecture is one YAML file, `<name>.archon.yml` — plain text executed by subagents, no runtime dependency. | field | meaning | | -------- | ------------------------------------------------------------------------------------------- | | `task` | `general` \| `reasoning` \| `coding` — sets generator style and judge emphasis | | `budget` | hard cap on subagent calls; planned calls are validated against it before execution | | `layers` | ordered list; each entry is `- component: {params}` — one component type per layer | Params: `count` (parallel modules in a generator or fuser layer), `samples` (repeats per generator), `top` (keep-K for ranker and unit_test_evaluator). ## Presets Used when no spec file is given; also the format by example. Scale `count` values p