I'm working on Maelstrom, an agent framework with only the basics:
- 'agent' as cognitive state, ie, what to think about
- 'workflow' as 'what to think about
- 'session' as immutable agent history
- 'timers' as a way to kick off an agent on a schedule (with or without a workflow attached
I've been working on this since just before OpenClaw dropped at the end of January. Currently it weighs in at around 20k lines of code. There is still a significant amount of work to be done on polish, but the core appears to be functional, and almost to the point where I can replace opencode as my daily driver (I'm very much looking forward to this).
From [1]:
---
I've been working on a framework since the end of January or so. I'm on my 7th draft. As I've gone along, each draft gets markedly smaller. The overlaps between what I'm building and openclaw are significant, but I've realized the elements that make up the system are distinct, small, and modular (by design).
There are only a few primitives:
1. session history
1a. context map + rendered context map (think of a drive partitioning scheme, but for context -- you can specify what goes into each block of context and this gets built before being sent out for inference).
2. agent definition / runtime
3. workflow definition / runtime
4. workflow history
5. runtime history (for all the stuff session and workflow history fail to capture because they are at a lower level in the stack)
That's it. Everything else builds on top of these primitives, including
- memory (a new context block that you add to a context map)
- tool usage (which is a set of hooks on inference return and can optionally send the output straight back for inference -- this is a special case inside the inference loop and so just lives there)
- anything to do with agent operating environment (this is an extension of workflows)
- anything to do with governance/provenance/security (this is an extension of either workflows and/or agent operating environment... I haven't nailed this down yet).
I suppose I should say something about how agents and workflows work together. I've broken up 'what to do' and 'how to think' into the two primitives of 'workflow' and 'agent' respectively. An agent's context map will have a section for system prompt and cognitive prompt, and an agent can 'bind' to a workflow. When bound, the agent has an additional field in their context map that spells out the workflow state the agent is in, the available tools, and state exit criteria. Ideally an agent can bind/unbind from a workflow at will, which means long-running workflows are durable beyond just agent activity. There's some nuance here in how session history from a workflow is stored, and I haven't figured that out yet.
Generally, the idea of a workflow allows you to do things like scheduled tasks, user UI, connectors to a variety of comms interfaces, tasks requiring specific outputs, etc. The primitive lays the foundation for a huge chunk of functionality that openclaw and others expose.
It's been fun reasoning through this, and I'll admit that I've had an awful lot of FOMO in the mean time, as I watch so many other harnesses come online. The majority of them look polished, and are well marketed (as far as AI hype marketing goes). But I've managed to stay the course so far.
I hope you find your ideal fit. These tools have the potential to be very powerful if we can manage to build them well enough.
---
[1] https://news.ycombinator.com/item?id=47784743