Personal exploration · private and unfinished
Exploring agent orchestration
A personal experiment in coordinating agents and approval steps to build applications and services. I am exploring how these workflows could run with local language models.
- Context
- Personal exploration
- My role
- Designer and developer
- Dates
- 2026 – present
- Status
- Private · unfinished · not on GitHub
Exploring
- Agent orchestration
- Approval workflows
- LLM-assisted development
- Local language models
Scope of this pageThis project is unfinished and private. It is not on GitHub or available as a product, and local-model support is still being explored. The diagram shows the workflow I am aiming for, not a tested result.
On this page
Why I'm exploring this
I enjoy finding new ways to solve problems. Agent orchestration connects that with system design: how to break a goal into work, coordinate agents, and decide when an output is ready to move forward.
What I am working on
I have been experimenting with a system that can create or adjust agents and coordinate their work on applications, with approval steps along the way. Some paths work already; the system as a whole is not finished and is not ready to publish.
The experiment builds on OpenClaw, an existing open-source agent platform. My work is the orchestration around the agents; I did not create OpenClaw.
The workflow I am working towards
- External service or data
- Decision
- Planned or proposed — not built
- Conceptual or illustrative
Text version of this diagram
Parts
- A goal — e.g. “build a small web service”; conceptual
- Break it into tasks — and suggest an agent for each; conceptual
- Approve the plan? — conceptual
- Create or adjust agents — conceptual
- Agents work in OpenClaw sessions — OpenClaw is an existing open-source platform; conceptual
- Review the output — checks and a person's decision; conceptual
- Next task, or done — conceptual
- Local language models — intended backend; support still in progress; planned, not built
Connections
- A goal → Break it into tasks (conceptual)
- Break it into tasks → Approve the plan? (conceptual)
- Approve the plan? → Break it into tasks: revise (conceptual)
- Approve the plan? → Create or adjust agents: approve (conceptual)
- Create or adjust agents → Agents work in OpenClaw sessions (conceptual)
- Agents work in OpenClaw sessions → Review the output (conceptual)
- Review the output → Create or adjust agents: changes needed (conceptual)
- Review the output → Next task, or done: accepted (conceptual)
- Local language models → Agents work in OpenClaw sessions: intended backend (planned, not built)
The questions guiding the work
- How should a larger request be divided into useful agent tasks?
- Where should a person review a plan or an output before work continues?
- What evidence should support a claim that a task is complete?
- How far can local models support these workflows, and where do their limits appear?
It also connects to my interest in responsible AI: building such a system raises practical questions about permission, oversight and accountability. I am treating it as a learning project, with no release date and no claim of production readiness.