The future of background agents

This post is 100% written by a human. (In spite of the subject matter)

I used to be a major AI skeptic. In fact, before June of last year, I was severely against tab complete. 6 months later I have many projects where I haven’t looked at a single line of code. It seems like since things have sped up rapidly in the past two months, in a way that we haven’t seen since GPT-3.

A big player behind this is OpenClaw - it’s the first time that I, and many others, have truly felt the power that AI will have over our lives. It’s a brilliant showcase of just how much benefit you can derive from LLMs. While I don’t believe it’ll stick around long-term, it has had a supremely profound impact on how we’ll further integrate AI into our lives.

The largest bottleneck I noticed when setting up OpenClaw was the single threaded nature of it. On a platform like Telegram, you can only have one ongoing conversation with it. That’s why I moved to Discord, allowing me to have as many channels as I wanted, but it still felt clunky; not designed from the ground up to run many sessions at once.

I began looking into how best to give OpenClaw better tools to be an orchestrator of agents. It has a sub-agent system, but it isn’t particularly powerful. I started off by allowing it to run and communicate with Ralph loops, and that worked great for a while, but they were two fundamentally different technologies that don’t play too well together.

Finally, we come to background agents. The team at Ona did a brilliant job of explaining them, but in a nutshell, they’re agents that are able to operate autonomously, without direct intervention from humans. The degree of their autonomy varies, but I like to classify them as agents that either begin or complete a task without our input. A perfect background agent would be one that can do both (correctly). That raises another question of how we can evaluate agent correctness, which is an arguably bigger topic.

Tying this back to OpenClaw, I built a system called ‘Dom’ (get it?) that allows it to build sub-agent workflows, then start, monitor, and evaluate them. This means that I can work with it to quickly spin up these agents, and leverage its stellar context and understanding of my repositories, without locking it up into a session. OpenClaw is also notoriously token-intensive, and delegating the work to a harness like OpenCode allows for a much more efficient use of tokens.

I might open-source Dom at same point, and do a deep dive into its architecture, but it’s still very much so a work-in-progress with a lot of sharp edges. If you have any thoughts around background agents, or agents in general, reach out!