BullX智能工作流

BullX — AgentOS for AI Colleagues

English | 简体中文 | 日本語

:warning: BullX is in early development. Some capabilities described here are still planned for later releases.

BullX is an AgentOS for working side by side with self-directed AI Colleagues.

Built on Elixir/OTP, PostgreSQL, and Redis, BullX is designed for durable digital work across enterprises, teams, and one-person companies.

OTP's lightweight processes, supervision trees, and message-passing isolation map naturally onto fleets of long-running, fault-tolerant AI Agents — see Why OTP is a better runtime for multi-agent orchestration for the longer argument.

Chatbots made LLMs conversational. The OpenClaw and Hermes-Agent generation gave agents hands: channels, tools, skills, shell and browser access, memory files, subagents, and scheduled work. Dify, RPA, and RAG workflow builders made it easier to package AI into specific business apps. BullX is designed for the next step: AI Colleagues that do accountable work over time, inside an operating model that can be audited, recovered, governed, and improved.

BullX is built around AI Agents as colleagues, not RAG support bots or digital assistants waiting for instructions. A BullX Agent can carry a long-term mission, track KPI/OKR-style success metrics, hold responsibility, work over long horizons, collaborate with humans or other Agents, and improve from trajectory data.

BullX does not optimize for "one more chat interface." It turns AI Colleagues into a durable work system:

• Agents carry long-term missions, responsibilities, permissions, memory, outbound identity, and KPI/OKR-style success metrics.
• IMGateway and other gateways emit CloudEvents mail and mirror or save the external-world facts they own.
• MailBox creates internal delivery entries for receivers such as AIAgents, Workflows, SubAgents, and gateways.
• Receivers do the work: most commonly an AIAgent for flexible judgment or a Workflow for explicit process structure.
• Principals, Budgets, and human collaboration paths make responsibility explicit before expensive or risky work happens.
• Capabilities expose models, tools, browsers, sandboxes, messaging channels, APIs, and external agent harnesses without hiding power inside prompts.
• Brain provides long-term memory and a reasoning world model: not a raw vector log, not a giant Markdown memory file, and not a fully predefined ontology, but evolvable knowledge extracted from conversations, events, actions, and outcomes.