Summary

AxcAgentEngine offers a Python framework for building AI agents, tackling task drift in complex scenarios by integrating Plan-Observe-Replan (POR) with the ReAct loop. This approach enables structured planning, dependency scheduling, and dynamic replanning, providing engineers and researchers with a more stable and predictable execution engine for sophisticated agent applications.

How It Works

The core innovation is the Plan-Observe-Replan (POR) strategy, augmenting the traditional ReAct loop. Agents first generate a structured plan, schedule steps by dependency, execute, observe, and replan. Managed by an Engine and extended via PluginRegistry, this allows complex task decomposition and error handling. Tools use ToolDefinition and must return ToolOutput, enabling concurrent reads and serial writes.

Quick Start & Requirements

Install via pip install axc-agent-engine. Requires Python and LLM API credentials (e.g., OpenAI). Supports various LLM providers. Examples demonstrate agent loading, chat, and streaming. Documentation covers Architecture, API, Plugins, and Examples.

Highlighted Details

• POR Planning: Structured plans with dependency scheduling and replanning (auto, react_only, por_first).
• Plugin System: Extensible via YAML-driven loading and explicit registration of built-in/custom plugins.
• Durable Recovery: CheckpointStore and ExecutionRecoveryService enable execution resumption.
• Advanced Memory & Knowledge: Four-layer memory (KV, dedup, decay, graph) with semantic chunking and hybrid retrieval (vector/BM25).
• Sidecar Suite: Specialized components for multi-agent sessions, simulations, evaluations, and cost analysis.
• OpenAI Compatible API: Offers a subset of the OpenAI Chat Completions API.

Maintenance & Community

The README does not detail specific contributors, sponsorships, or community channels. Contributing and security guidelines are available via documentation links.

Licensing & Compatibility

Licensed under Apache-2.0, permitting commercial use and integration into closed-source applications.

Limitations & Caveats

The API intentionally omits request-level tool_choice and n > 1. Users should verify /v1/capabilities for full OpenAI parity. LLM configurations are in host code, not agent YAML. All tools must return ToolOutput.