Single source of truth for every tunable lever of the deep audit-question agent (loop bounds,
toolset, prompt fragments). Model and temperature come from the LangSmith hub / committed
snapshots (with optional per-prompt/agent constructor overrides) — not from this config.
The agent stays pure — it accepts a (partial) config and resolves it against these defaults.
Production callers resolve the partial from LaunchDarkly at the call site and pass it as
input.config, so @repo/ai-agents never takes a @repo/feature-flags dependency.
Single source of truth for every tunable lever of the deep audit-question agent (loop bounds, toolset, prompt fragments). Model and temperature come from the LangSmith hub / committed snapshots (with optional per-prompt/agent constructor overrides) — not from this config.
The agent stays pure — it accepts a (partial) config and resolves it against these defaults. Production callers resolve the partial from LaunchDarkly at the call site and pass it as
input.config, so@repo/ai-agentsnever takes a@repo/feature-flagsdependency.