Making Agent Skills Disposable: Workflow Logic Should Live Below the Prompt

jonah_omninode · reddit · 2026-08-25

The author argues most agent skills are instruction bundles re-interpreted on every run; putting workflow logic inside them means CLI, MCP, hook, and direct-run paths become divergent implementations of the same idea.

OmniNode is moving to a thinner boundary: surface adapter -> typed request -> standard runner -> node/handler -> typed result -> evidence receipt. The skill owns argument parsing, typed input construction, invocation, and result rendering — not the workflow logic. The underlying package carries the contract, handlers, tests, a recorded success path, typed failure behavior, and protocol bindings; local installs use in-memory transport while self-hosted deployments swap adapters.

Core principle: if a rule matters enough to reject work, it belongs in a schema, validator, permission boundary, or gate. The prompt explains intent but should never be the only thing preventing an invalid state. Honest limitation: only one database seam has been proven across local, production, and test bindings.

Original post →

More from coding & agent

coding & agent channel →