# Agent Zero (AI) Agent Zero is an [[Open Source|open-source]] agentic framework that gives an [[AI Agents|AI agent]] a complete, real computer to work in rather than a sandboxed set of API calls. The whole thing runs inside a [[Docker]] container with a full Linux environment — including an XFCE desktop — so the agent can drive actual GUI software (Blender, LibreOffice, a browser, a terminal) the same way a human would. The defining idea is **transparency through prompts-as-code**: behaviour isn't buried in framework internals, it lives in human-readable, editable prompt and config files. You change what the agent does by editing its prompts, tools, and extensions — not by patching source. That makes it closer to a hackable operating system for agents than a fixed product. ## How it works - **Dynamic multi-agent delegation.** Any agent can spawn subordinate agents to handle sub-tasks. Each subagent keeps a focused context and reports findings back up, which keeps the top-level context clean. There's no fixed org chart — the hierarchy forms on demand. - **Tools, skills, extensions.** Custom tools live in `/tools/` or community plugins; skills are dynamically-loaded expertise modules (SKILL.md style) that only enter context when relevant; extensions are [[Python]] hooks into the agent lifecycle for deeper behavioural changes. - **Memory.** Hybrid memory built on FAISS vector search, split into main memories, conversation fragments, and proven solutions, with consolidation strategies. - **Interoperability.** Speaks [[Model Context Protocol (MCP)]] as both client and server, and Agent-to-Agent (A2A) via FastA2A so multiple instances can cooperate. Privacy-respecting web search via SearXNG. - **Coworking.** Browser "annotate mode" and a shared Markdown/LibreOffice editor treat human and agent edits as equal operations. Snapshot history gives time-travel: inspect, diff, and roll back the workspace. ## Why it matters Most agent frameworks abstract the environment away and hide their reasoning. Agent Zero inverts both: the environment is a real Linux box the agent fully controls, and the reasoning is exposed prompt text you own and edit. The tradeoff is responsibility — a transparent, general-purpose computer-using agent is powerful precisely because so little is locked down. Built by the `agent0ai` organisation; primarily [[Python]] with a [[JavaScript]]/HTML Web UI. ## References - https://github.com/agent0ai/agent-zero - https://www.agent-zero.ai/ - https://www.agent-zero.ai/p/docs/ - https://www.agent-zero.ai/p/architecture/ - https://www.agent-zero.ai/p/use-cases/ - https://www.agent-zero.ai/p/community/ ## Related - [[Artificial Intelligence (AI)]] - [[AI Agents]] - [[Docker]] - [[Open Source]]