wavefront

Abstracts multiple LLM providers (OpenAI, Anthropic, local models via Ollama) behind a unified API layer, enabling seamless model swapping and provider-agnostic agent development. Routes requests through a provider registry pattern that handles authentication, rate limiting, and response normalization across heterogeneous APIs without requiring application-level conditional logic.

Coordinates multi-step agent execution by managing tool/function calling, state transitions, and decision branching through a declarative workflow definition. Integrates with CrewAI and LangGraph patterns to handle agent-to-agent communication, tool result injection, and loop termination conditions without manual state management.

Tracks LLM usage costs by monitoring token counts, API calls, and provider-specific pricing models. Integrates with billing systems to generate cost reports, set spending limits, and allocate costs across projects or teams. Supports real-time cost alerts and cost optimization recommendations.

Manages end-to-end RAG workflows including document ingestion, chunking, embedding generation, vector storage, and semantic retrieval. Supports multiple embedding models and vector databases (Pinecone, Weaviate, local FAISS) through a pluggable backend architecture, with built-in reranking and context window optimization.

Enforces content safety, prompt injection detection, and output validation through a policy-based filtering system. Integrates with guardrail frameworks (e.g., Guardrails AI) to apply rules before LLM calls and after generation, supporting custom validators, PII masking, and jailbreak detection without modifying agent code.

Captures detailed execution traces of agent workflows including LLM calls, tool invocations, latency metrics, and error states. Exports traces to observability platforms (Langfuse, LangSmith) or local storage in structured JSON format, enabling debugging, performance analysis, and audit trails without instrumenting agent code.

Provides a Python framework for building MCP servers that expose tools, resources, and prompts as standardized protocol endpoints. Handles MCP protocol serialization, request routing, and error handling, enabling agents to discover and invoke capabilities across process boundaries using standard MCP client libraries.

Packages agents and middleware components as Docker containers with built-in health checks, graceful shutdown, and resource limits. Supports Kubernetes deployment with service discovery, load balancing, and horizontal scaling of stateless agent instances without requiring manual orchestration configuration.

Enables agents to discover available tools at runtime by querying tool registries and validating tool invocations against JSON Schema definitions. Supports automatic schema generation from Python function signatures, runtime schema validation, and type coercion to prevent invalid tool calls.

Automatically manages LLM context windows by chunking long documents, summarizing irrelevant sections, and prioritizing recent/relevant information. Uses sliding window techniques and importance scoring to fit maximum relevant context within token limits while preserving semantic coherence.

Manages conversation state across multiple turns by maintaining message history, user context, and agent state in a persistent store. Supports session resumption, context recovery, and conversation branching for exploring alternative agent responses without losing conversation history.