Understand model preparation¶
Before a run, Soulacy examines the selected provider/model and uses available capability metadata to choose a compatible working approach. It adapts the method, not the promise you asked the agent to fulfill.
Inspect it before running¶
- Open the agent in Agents (web or iPhone).
- Save any model/configuration edits first.
- Open Model preparation and choose Refresh preparation.
- Read the provider/model, strategy, evidence source, warnings, and any blocking reason. Expand Capabilities and working approach on the web.
The preview reads saved configuration. Metadata may be cached for up to five minutes; refreshing the panel is not a fresh model benchmark.
Interpret the result¶
| What you see | Meaning | Your next step |
|---|---|---|
| Native tools supported | Provider metadata reports tool-call support | Test the task; support does not guarantee correct tool arguments |
| Tools known unsupported, automatic strategy | A guarded JSON/ReAct approach may be used | Test with read-only tools before considering writes |
| Capabilities unknown/not reported | The provider did not supply usable metadata | Do not assume incapable or highly capable; run a small representative test |
| Input/output/context ceiling | Reported/configured capacity, when available | Keep the task and history within it; absence is not unlimited capacity |
| Blocking reason | A requirement is incompatible with known capability/configuration | Deliberately choose a compatible model or change the requirement yourself |
It never invents an intelligence score from a model's name, silently drops required output, grants tools, expands your budget, downloads a model, or switches to a paid model on its own.
A useful comparison exercise¶
Run the notes exercise on a model you already have, keeping the prompt and acceptance checks fixed. Record errors, duration, and quality. If you deliberately choose a second model, save it, inspect preparation again, and repeat the same normal and failure cases.
Do not equate a larger context window or reported reasoning support with better answers. If the unchanged task cannot fit a model's limits, stopping is safer than silently discarding required context.
The technical reference covers providers, cache behavior, explicit strategies, nested budgets, API errors, and SDK integration.