AI architecture: a multi-agent ensemble at single-FTE scale

Forum Atlas's AI is not a single Claude call. It's a coordinator agent orchestrating seven specialists plus a citation validator. Single-provider Anthropic — Sonnet 4.6 as the workhorse, Haiku 4.5 for volume + validator, Opus 4.6 for high-stakes (Forecast).

Why single-provider at v0? Because multi-provider routing adds engineering + evals + drift management cost that single-FTE Phase 0 can't absorb. The trigger to add a second provider is $1M+ ARR or a sustained Anthropic outage breaching customer-committed SLA — neither of which has happened.

Embeddings are voyage-3 at 1024 dimensions. Why not OpenAI text-embedding-3-small? Voyage-3 leads MTEB-Government benchmarks for long-form regulatory text, which is the corpus shape Forum Atlas operates over. Reranker is Cohere Rerank v3.5, chosen to lift retrieval precision over cosine-only ranking.