axclogo/AxcAgentEngine

What it solves

Most agent frameworks rely on a simple "think-act-observe" loop (ReAct), which can drift or lose focus as tasks become more complex. AxcAgentEngine solves this by introducing a Plan-Observe-Replan (POR) mechanism, allowing agents to create structured plans, schedule steps based on dependencies, and adapt their strategy based on outcomes.

How it works

The engine combines a standard ReAct executor with a POR planning layer. It uses a plugin-based architecture where core capabilities—such as memory, knowledge retrieval, and risk guarding—are decoupled from the engine and loaded via YAML configurations. It separates stable agent behavior (defined in YAML) from runtime resources (like LLM providers and database indexes), which are bound during instantiation. To handle large data, it uses an ArtifactStore to manage oversized tool results without bloating the LLM's context window.

Who it’s for

Developers building complex AI agents that require structured planning, durable workflows, and enterprise-grade controls like human-in-the-loop approvals and risk classification.

Highlights

• POR Planning: Supports structured plans and dependency scheduling to prevent task drift.
• Extensible Plugin System: Modular capabilities including memory (four-layer), knowledge (hybrid retrieval), and safety (PII masking).
• Durable Workflows: Includes a WorkflowRuntime and CheckpointStore to allow agents to resume from previous states.
• MCP Support: Compatible with Model Context Protocol (MCP) via stdio and JSON-RPC HTTP.
• Sidecar Suite: Specialized tools for multi-agent orchestration, simulation, evaluation, and failure mining.
• OpenAI Compatibility: Provides an OpenAI-compatible HTTP client and API subset.