What would it look like if your AI agent could just… work? Not need babysitting, not hallucinate mid-task, not forget everything the moment you close the tab? That’s the question we spent all of 2025 trying to answer.

Xaibo - The Framework for Production AI Agents

We released Xaibo in June to solve what we call the “tutorial cliff” - the moment an agent framework’s simplicity becomes an uncontrollable black box. (We wrote about the engineering lessons behind this in Operationalizing AI Agents.) Xaibo is designed specifically for production deployments, around three core principles:

• Protocol-driven design - agents use components that communicate through well-defined interfaces, configured in simple YAML. Swapping LLM providers or databases becomes trivial.
• Dependency injection - every component explicitly declares what it needs, making testing straightforward and configuration explicit.
• Transparent observability - every interaction is automatically captured and visible, eliminating guesswork when debugging.

The result is a framework where simple cases stay simple, but complex production requirements become manageable instead of impossible.

Check out the Xaibo Manifesto for the full technical story, or grab the code on GitHub. Xaibo became the foundation for everything that followed.

XpressAI OS: Running Agents at Scale

With the framework in place, we started building XpressAI OS - our platform for running Xaibo agents 24/7 at enterprise scale. The vision was simple: we wanted to make the agents that work alongside our users like coworkers instead of chatbots.

When you sign up for XpressAI, you’re hiring your first digital team member. Choose from templates like Sales Development Rep, Content Marketing Manager, DevOps Engineer, or Customer Success Agent. Each comes pre-configured with the skills and workflows for their role. Want a Sales Development Rep? They’ll qualify leads, send personalized outreach, and schedule meetings. Need a Customer Success Agent? They’ll handle support tickets, send proactive updates, and maintain your knowledge base.

For agents that need access to a computer, XpressAI OS provides full computer use support. Each agent gets its own isolated virtual desktop where it can control a mouse, type on a keyboard, and navigate applications like a human would. This is especially valuable for tasks that are destructive or risky to run on your own machine - let the agent browse untrusted websites, test suspicious software, or run scripts in a sandboxed environment where nothing can touch your actual system. It’s also the answer for legacy software, internal tools, and anything that was never designed to be automated through an API.

We also made sure it fits into our user’s existing workflow. Xaibo agents can live in Slack channels alongside the human team and share the same context. The platform handles budget controls, shared knowledge bases, human-in-the-loop workflows, and on-premise deployment for regulated industries — and it supports the full model stack: OpenAI, Anthropic, Google, Mistral, and local models like Llama and Gemma for teams that need complete control. (We learned the hard way why this matters — see our agent lessons post.)

XpressAI CLI: The Open Source Agent Runtime (Sneak Peek)

Alongside the commercial platform, we’re preparing to release XpressAI CLI - our take on the now-popular OpenClaw. It’s an open-source “operating system for AI agents” built around radical simplicity: xpressai init to create a workspace, then xpressai up to start your agents. Zero configuration to get started, but progressive disclosure for power users who need fine-grained control.

We designed XpressAI CLI specifically for autonomous agents:

• Memory - A three-tier system (near-term slots, Zettelkasten knowledge graph, and vector search) with hooks that automatically recall context and extract important information. This solves the “amnesia problem” where agents forget everything between sessions.
• Budget controls - Tracks daily, monthly, and per-task spending with configurable enforcement: pause, degrade to a cheaper model, alert, or stop.
• Tooling - Uses MCP (Model Context Protocol) as the universal standard. Built on Xaibo’s architecture, so you get full observability into every tool call, memory retrieval, and decision.
• Backends - Claude, OpenAI, Google, and local models like Qwen3-8B.
• Web UI - While it’s primarily a CLI, it also ships with a built-in web app for when you’d rather not live in the terminal. Manage agents, monitor tasks, and track budgets from the browser.

We’re still building on it, but it’s already shaping up to be useful for running agents continuously on development machines or small-scale production deployments. The code will be on GitHub soon, and we’re actively developing scheduled tasks, multi-agent orchestration, and better debugging tools for long-running behaviors.

What’s Next

We want to enable the first single-person unicorn powered by a digital workforce. We’re getting closer - here’s what’s coming:

• XpressAI OS beta - We’re opening it up soon. Join the waitlist at xpress.ai to be among the first to run agents that work like real team members.
• Bot/Agent Store expansion - More turnkey templates so you can deploy a useful agent on day one: calendar managers, meeting summarizers, invoice processors.
• XpressAI CLI public release - For developers and hobbyists who want agents running continuously on their own machines, not just responding to pings.
• Fine-tuning in Xaibo - Once you’ve gathered real usage data from your workflows, train models on how your business actually operates. That’s when agents become true assets - uniquely trained for your needs, not just generic tools.

Build something real. See what agents can actually do.

Want to try XpressAI? We’re opening up beta access soon - join the waitlist at xpress.ai or check out the open source framework at xaibo.ai. Join our Discord to connect with other builders deploying agents in production.