LLM provider freedom

LIFEOSAI runs on any LLM provider the harness can reach. The orchestration substrate is provider-agnostic — each agent picks its model in its adapter_config.

Supported providers

Provider	Models	Best for
Anthropic via GCP Vertex AI	Claude Opus 4.6, Sonnet 4.6, Haiku 4.5	Default for NBS workshop. Highest reasoning quality.
Anthropic direct	Same	Slightly lower latency; direct API
OpenAI	GPT-4, GPT-5 series	When OpenAI-specific tools or strengths matter
Server-hosted OSS-80B	Llama-3.3-70B, Qwen-2.5-72B, etc.	Cost-conscious workloads where frontier quality isn’t required

Per-agent model choice

Different agents need different models. Examples from our running config:

Incident Collector — Sonnet (search + normalize; doesn’t need Opus’s depth)
Root Cause analysis — Opus (high-stakes causal reasoning)
Threat Modeling — Opus or Sonnet depending on incident complexity
Critic — Opus (judgment-heavy)
Document converter (Stage 2 in library) — Haiku (cheap + fast)

This mix-by-role keeps quality high where it matters and cost low where it doesn’t.

GCP Vertex AI is the workshop default

For NBS the path is Claude → GCP Vertex AI → your NBS GCP project. Configuration in Setup:

CLAUDE_CODE_USE_VERTEX=1
ANTHROPIC_VERTEX_PROJECT_ID=<your-NBS-project>
CLOUD_ML_REGION=asia-northeast1 (Tokyo) or your preferred region
ANTHROPIC_MODEL=claude-opus-4-6@... etc.

Same env vars work for both Claude Code (the harness) and LIFEOSAI (which calls Claude Code internally).

Cost positioning

Cost-consciousness is a common enterprise requirement. LIFEOSAI’s provider-agnostic design supports this directly:

Mix high-cost (Opus) with low-cost (Sonnet, Haiku) per agent role
Plug in OSS-80B for high-volume workloads
All without changing the orchestration code — just adapter_config