Agent workflow

Sol-Engine is agent-native: a coding agent (Claude Code / OpenAI Codex) drives model onboarding and optimization, not just fixed scripts. This page documents the orchestration behind the "hand an agent a goal" quick start.

Orchestration modes

Lightweight (master + executors). A single master orchestrator supervises up to three executor sub-agents, one per acceleration technique (kernel, cache, sparse/PISA). Each executor explores its technique's search space (detached, watchdog-guarded), returns validated candidates, and the master gates, dedupes, and integrates them. Preferred for a fast, bounded optimization sweep on a new model.

Full framework. Per-technique executor nodes with richer fan-out for deep multi-round searches.

Candidate contract

Every optimization is a candidate manifest: an id, kind (baseline / patch / control), the runtime it uses, the environment flags that turn the technique on, and its eval profile. A single launcher renders a reproducible run bundle and submits it. Baselines and same-topology controls are first-class, so every speedup has an explicit, stated denominator (single-GPU / same-topology / vs-naive).

Quality gates

No speedup is accepted on wall-clock alone:

• LPIPS against frozen baseline frames.
• Multimodal visual gate — a hosted VLM reviews baseline-vs-candidate side-by-side against a rubric (snow/speckle, blur, mosaic / patch-boundary, banding, ghosting, melting, temporal flicker, coherence & motion regressions).
• Authenticity check — confirms the technique actually engaged (PISA wrote its stats, cache actually reused steps) so a run cannot report a "fake" speedup from a no-op optimization.

Conventions

• Report speedups against a stated baseline; never mix denominators silently.
• Timing excludes model load; medians over official validation prompts; same seed.
• Optimizations that change floating-point reduction or attention sparsity are marked non–bit-exact.