Interesting approach. One pattern I've seen work well with YAML-defined agent workflows is keeping your session state observable alongside the orchestration config. When agents run in parallel, the hardest debugging problem isn't the workflow definition itself but correlating what each agent actually did across concurrent runs. Whatever tooling you use, prioritize being able to replay and search past agent output by context, not just timestamps. That audit trail saves hours.
Actually, Kelos captures structured outputs after each coding task compeletes: the final branch, the created PR link if exists, input/output tokens, and so on. And coding agent’s logs would remain after finished as a container log.
That sounds solid. The structured outputs per task are useful, especially the token tracking.
One thing worth considering: a lightweight timeline view that correlates those per-task outputs across parallel agents. When something goes wrong mid-workflow, jumping between container logs to reconstruct what happened concurrently gets tedious fast. Even a simple ordered event stream tying the branches and PR links together would save a lot of debugging time.
Interesting approach. One pattern I've seen work well with YAML-defined agent workflows is keeping your session state observable alongside the orchestration config. When agents run in parallel, the hardest debugging problem isn't the workflow definition itself but correlating what each agent actually did across concurrent runs. Whatever tooling you use, prioritize being able to replay and search past agent output by context, not just timestamps. That audit trail saves hours.
Thanks,
Actually, Kelos captures structured outputs after each coding task compeletes: the final branch, the created PR link if exists, input/output tokens, and so on. And coding agent’s logs would remain after finished as a container log.
Is theere anything you want to add additionally?
That sounds solid. The structured outputs per task are useful, especially the token tracking.
One thing worth considering: a lightweight timeline view that correlates those per-task outputs across parallel agents. When something goes wrong mid-workflow, jumping between container logs to reconstruct what happened concurrently gets tedious fast. Even a simple ordered event stream tying the branches and PR links together would save a lot of debugging time.