1. chatting with Claud Code about the general plan
2. assigning slices of work to other worker agents like Pi + DeepSeek 4.1 Flash
3. building in parallel
4. Claud Code reviews when done, runs tests, and merges after I validate
It worked, but with a lot of glue in the middle, so I built radian + arc into the same binary.
Radian: a compact, pragmatic agent that has enough tooling to solve a small coding task end to end; it does not try to be a plugin platform. You can still extend it via MCP, skills, or hooks to extend its capabilities. I think plugins are great, but a lot of them are for humans. I'm trying to remove humans from the loop. Humans are at the start and end of the loop.
arc: built into the radian binary;
* it can get work sent from another agent
* assign work items to concurrent headless radian agents
* automatically review, CICD, integrate locally
* humans can still be in the loop but not at the agent conversation or code review level, but at the work item approval/review level.
* shows all of the above in a plan view or Kanban board in TUI.
I have been heavily dogfooding, and Radian + arc basically built themselves in the last couple of weeks, with me just being a product manager/scrum master and qa.
radian + arc are open source at https://github.com/hisea/radian
What do you think about this approach? Looking for ideas to make Radian better or fit more scenarios.