shibing624/agentica
Build AI agents that run for hours, not seconds. Async-first Python harness: auto context-compaction, cost budgets & death-spiral guards, persistent cross-session memory, self-evolving SKILL.md, /goal long-task loop, 40+ tools, 20+ models, MCP/ACP. Ships a CLI + Web Gateway out of the box.
What it solves
Standard LLM interfaces often struggle with long-running tasks, infinite loops, and losing context over time. Agentica provides a robust execution engine designed to keep agents running reliably without "flying off the rails," managing costs, and preventing infinite loops during complex tool-use cycles.
How it works
The framework uses an async-first agentic loop driven by a control-flow engine. It manages long-running goals through a specialized /goal loop that automatically determines when a task is complete or needs to pause. It handles memory through indexed storage and relevance-based retrieval, and features a unique self-evolution pipeline where successful task sequences are compiled into reusable Markdown-based SKILL.md files.
Who it’s for
Developers building complex AI agents that require more than just simple chat interfaces, specifically those needing multi-agent collaboration, persistent cross-session memory, and autonomous tool execution (like web searching, file manipulation, or code execution).
Highlights
- Self-Evolving Skills: Automatically compiles successful execution sequences into reusable skills.
- Long-Task Management: Supports
/goalloops with hard caps on tokens, time, or turns. - Robust Execution: Built-in context compression, cost budgeting, and infinite loop protection.
- Extensive Integration: Supports 20+ models and 40+ built-in tools including MCP and RAG via LangChain/LlamaIndex.
- Flexible Deployment: Includes a CLI, a Web Gateway, and integration for IM platforms like Telegram and Discord.