I'm exploring the concept of self-awareness in locally stationed Claw-style agents using a multi-layer agent orchestration kernel. My hope is that by using novel local short-term and long-term memory architectures, we can improve current context-engineering harnesses to gain consistent self-reflection, self-correction, and self-improvement over indefinite periods of time.

The fall-short failure mode of this project is a very competent personal assistant agent that is built for security and auditability. The goal is to minimize misinterpretation (aka rogue actions) through hard permission gating, and transparent process records via immutable durable logs.

To help make the dream of democratized and soveriegn personal agents a reality, I am building all of this open source, with a supporting package management platform and optional hosting and token packages, if you want to help support the business.

I'm building this with a multi-tier agent swarm. A PM agent coordinates the main thread with six project workflows at hand (state audits, codebase audits, ideation, gap analysis, ticket creation, sprint composition), alongside persistent workflow designer and architecture discovery agents. Execution fans out across 2-6 parallel lanes, each a 14-agent research-and-engineering pipeline with fresh reviewers spun up per gate. Concurrent agent count peaks around 39 with 6 lanes active, climbing to 57 when cleanup and close workers fire. Currently achieving a 6× reduction in token burn through progressive disclosure, routing tables, and handoff-artifact-based context management.

Contributions welcome — check out the open issues or read the contributing guide.

Find out more at orthg.nl.