Automatic learning
How an agent reviews its work and preserves useful lessons without retraining a model.
Jian's learning is a review loop over completed work. The agent can retain a durable fact as a memory or a repeatable method as a skill. It does not fine-tune the provider model, change model weights, or guarantee that every correction will be retained correctly.
The useful outcome is inspectable text you can read and change, not an invisible new model version.
When a review is considered
In the 1.1 implementation, a completed normal turn qualifies when it:
- Used at least 10 tools.
- Encountered a failed tool before reaching completion.
- Reached the periodic threshold of 15 completed turns since the previous review.
There is a 10-minute cooldown per profile. Learning reviews do not recursively trigger more learning reviews. These are runtime thresholds, not a promise that a review runs after every correction or every fifteenth visible message.
If Decisions/Jev is available for learning, the gateway first asks whether a short account of the work contains something worth keeping. A low usefulness score can skip the full review. Without that judgment, the fixed trigger rules still work.
What the review sees
The gateway constructs a bounded brief from the request, tool names and failures, and the answer. Periodic reviews include excerpts of recent completed turns. This is not an unrestricted rereading of every conversation on the installation.
The review runs in the profile's Learning conversation, created on first use. Its response says what was kept and why, or that there was nothing to keep. Use that conversation and the resulting Memories/Skills entries to inspect the effect.
A deliberately narrow tool set
Learning has at most eight execution steps, further bounded by the profile's policy. Its available tools are limited to:
- Reading, writing, deleting, linking, and unlinking memories.
- Creating, updating, and deleting the agent's own skills.
- Loading a skill's instructions.
It is not given shell, Computer Use, messaging, or MCP tools for the review. A background look-back should not become an unrequested external action. Normal agent work and learning are separate execution contexts.
What should become a lesson
| Experience | Useful retained form |
|---|---|
| “Lead my updates with the result, not the chronology.” | A standing preference in memory |
| A verified sequence for preparing a recurring report | A skill with the steps and checks |
| A command failed because a project requires a specific working directory | A precise workflow correction, if it is reusable |
| A one-off network timeout | Usually nothing |
| A secret visible in an error | Nothing; secrets do not belong in memory or skills |
A review can be wrong. Inspect generated entries, correct their scope, and remove a procedure that overgeneralizes one task. Agent-written skills cannot overwrite owner-written or imported skills through the agent's skill-editing tools.
Controls and cost
Learning is part of the current agent runtime. The Decided by Jev → Learning setting controls whether Jev helps decide which work deserves a review. Turning that setting off switches back to the fixed rule; it is not a global off switch for learning.
Reviews make model calls and therefore consume provider usage. The cooldown, bounded brief, and short step budget limit overhead; they do not make it free at the model provider. Disabling a particular built-in skill also does not imply that the runtime review scheduler is disabled.
Make corrections actionable
Instead of “that was wrong,” identify the rule and where it applies:
For Harbor Studio's release notes, describe the user-visible change first. Keep internal file names out unless they explain a compatibility issue. Use that format for future release notes.
The agent can act on a correction in the current conversation, and may preserve it immediately through its ordinary memory or skill tools. The background review is an additional opportunity, not the only way it learns.
Implementation reference
Review triggers and brief, runtime tool restrictions, and Decisions fallbacks.