Schedulers and workflow engines often expose human-oriented CLIs with flags and prose output. Agents work better with:
- raw JSON input
- schema discovery
- narrow field selection
- stable machine-readable output
- protocol surfaces instead of screen scraping
agentcli sits above execution runtimes.
- It can work standalone for authoring, validation, planning, and JSON-RPC.
- It can target
openclaw-schedulerfor durable execution and inspection.
- keep the manifest portable
- keep output structured
- prefer explicit schemas over docs-only conventions
- avoid duplicating runtime durability in this repo