UAP Platform
The Unified Agentic Platform — a governed runtime and SDK that lets teams build, deploy, and operate autonomous agents safely, in hours instead of months.
UAP is the foundation our products are built on. It packages everything an autonomous agent needs — a base agent framework, model routing, retrieval, session and memory, messaging, and a full lifecycle and governance layer — into one pip-installable SDK (platform_core). Product teams subclass a base agent and register their own tools and policies, shipping domain agents without re-implementing the plumbing, while platform-level guardrails, tracing, and approvals keep every action safe and auditable. It runs on your own infrastructure — no data leaves your cloud.
What it does
Base agent framework
A reusable LangGraph agent — Scanner/Input → Retriever → Reasoner → PolicyGate → Executor → ResponseAssembler — that products extend in a few lines.
Model gateway & routing
A LiteLLM gateway routes to Claude Sonnet (primary), Haiku (fast), and GPT-4o (fallback), with retries and failover, and traces every call to LangSmith.
Deterministic guardrails
Six non-LLM evaluators — PolicyCeiling, AuthRequired, ConflictCheck, PIIFence, EscalationTrigger, TenantPermission — return pass / warn / block before any action.
RAG & data layer
Postgres + pgvector hybrid retrieval, a Redis session store, and a run/trace schema that makes every agent run observable.
Governance & multi-tenancy
Catalog, versioning, promotion, and certification, with row-level tenant isolation so many agents can serve many tenants safely.
Fine-tune flywheel
LangSmith export → Together.ai fine-tune → vLLM shadow mode, so agents improve from their own traces before anything is promoted.
- 1
Plan
The agent decomposes the task into a traceable plan.
- 2
Retrieve
Hybrid pgvector + full-text retrieval grounds it in your data.
- 3
Reason
The gateway routes to the right model with retries and fallback.
- 4
Guard
Deterministic evaluators gate every proposed action.
- 5
Act / approve
Low-risk actions execute; high-risk pause for a human.
Built and merged; control plane in rollout.
Feature-complete foundation: the gateway, data layer, agent framework, guardrails, tenancy, lifecycle governance, and the fine-tune flywheel are built and merged. The self-serve control plane — the lifecycle and tenant-onboarding API and console — is deployed and in rollout. Next: OIDC/RBAC via IAM Identity Center and broader multi-cloud data and messaging.
Good questions.
Is UAP a framework or a managed platform?
Both. platform_core is a pip-installable SDK your teams build on, running on your own infrastructure (EKS/IRSA), with a control plane for lifecycle and tenants. No data leaves your cloud.
How is it different from using an agent library directly?
A library gives you an agent loop. UAP adds the parts that make agents safe in production — deterministic guardrails, human approvals, tracing, versioning, and multi-tenant governance.
Which models does it use?
Claude Sonnet as primary and Haiku for fast routing, with a GPT-4o fallback — all behind a gateway, so models are swappable and every call is traced.