Learning step
A named, whitelisted bucket of rules an agent reads, where corrections from past runs accumulate instead of turning into scattered prompt edits.
Last verified
A learning step is a named bucket of rules that one or more agents read, where "we learned to always do X" accumulates in one whitelisted place instead of turning into ad hoc edits scattered across system prompts.
How it's structured
A learning step is authored at learnings/<name>.yaml with a title, a
description of what kind of rule belongs there, an optional preamble for
context, and the list of agents that own or read it. The individual rules
inside a step are runtime rows, written through the dashboard as they're
learned — the workspace file only declares the bucket and who can use it.
Why it matters in production
Agents make mistakes that get corrected in the moment, and without a structured place to keep the correction, that knowledge either lives only in whoever caught the mistake or gets bolted onto a system prompt as one more paragraph nobody reads carefully. A named, whitelisted step means a correction is scoped to a specific kind of judgment call, attributed to the agents that need it, and reviewable the same way any other workspace file is.
How Vocion implements it
LearningStepManifestSchema
(packages/core/src/libs/workspace/schemas.ts) validates the file, which
applies to the learning_step table; individual rules attach as learning
rows at runtime and render into the agent's virtual filesystem at
/learnings/<name>.md. Full field reference:
docs/entities/learning-step.md.
The dashboard surface is /dashboard/learnings.