I'm the CTO of Kolega and the main author of this. Kolega Code is an open source terminal coding agent. What I want to show today is Gigacode, its orchestration engine. The model writes its own multi-agent workflow as a python program and the harness executes it.
When Claude Code shipped dynamic workflows under the ultracode name a few months ago I thought it was the most interesting thing that happened to coding agents for a while, but it's closed source. This is the same idea, except you can read the code. The model receives your request, maybe investigates a bit, then emits a program written against a small runtime API with paralle phases, pipelines, judge panels, etc.
Architecture doc, including what's guaranteed by the runtime, the journaling and resume behaviour, and what workflow shapes the models actually chose in practice, drawn from a few hundred of our internal runs: https://github.com/kolega-ai/kolega-code/blob/main/how-gigac...
Main thing that needs improvement: the workflow current runs in the foreground and blocks the chat thread until it finishes. Backgrounding it is the next piece of work.
What tasks are you working on with one serial agent that a parallel workflow or a swarm might do better? Interested to see what people might apply this to.
Hi HN,
I'm the CTO of Kolega and the main author of this. Kolega Code is an open source terminal coding agent. What I want to show today is Gigacode, its orchestration engine. The model writes its own multi-agent workflow as a python program and the harness executes it.
When Claude Code shipped dynamic workflows under the ultracode name a few months ago I thought it was the most interesting thing that happened to coding agents for a while, but it's closed source. This is the same idea, except you can read the code. The model receives your request, maybe investigates a bit, then emits a program written against a small runtime API with paralle phases, pipelines, judge panels, etc.
There's an unedited workflow example the model authored in the repo: https://github.com/kolega-ai/kolega-code/blob/main/examples/...
Architecture doc, including what's guaranteed by the runtime, the journaling and resume behaviour, and what workflow shapes the models actually chose in practice, drawn from a few hundred of our internal runs: https://github.com/kolega-ai/kolega-code/blob/main/how-gigac...
Main thing that needs improvement: the workflow current runs in the foreground and blocks the chat thread until it finishes. Backgrounding it is the next piece of work.
What tasks are you working on with one serial agent that a parallel workflow or a swarm might do better? Interested to see what people might apply this to.