Register any model via OpenAI-compatible API: vendor-hosted (OpenAI, Anthropic, etc.) or self-hosted (Nemotron, Llama, Mistral, etc.). Keys managed securely. Metadata (context windows, capabilities, internal/external classification) tracked per model.

Register Model Context Protocol servers (internal or external) with automatic tool discovery. Agents see only tools they're authorized to access.

Define agents using Oracle's AgentSpec DSL (JSON). Platform compiles to the configured runtime. Agents are version-controlled with full lineage tracking.

Every action against the platform passes through Cerbos-based policy enforcement. Policies define which models & tools agents can access, how internal data flows to external providers and what actions are authorized across the agent graph. All rules are configurable via YAML, enabling flexible, environment-specific governance without hardcoding logic.

Every agent runs instrumented end-to-end via OpenTelemetry. Full audit logs with literal input/output content at every step.

Based on MemGPT research. Treats LLM as an operating system. Agent decides when to access memory. Enables small models (Claude Haiku, GPT-4o mini) to handle deep, multi-turn research by automatically managing context window economics. Creates knowledge graphs during execution.

Set limits at any level: per model, per team, per month. Control token budgets, tool call limits, delegation depth. Prevent runaway costs and infinite agent recursion.

Configurable profiling integration provides flame-graph style execution timelines. See where time is spent, identify bottlenecks, auto-tune hyperparameters against evaluations.