TASIP-006: Cross-deployment learning

[ryw]

Status: Draft — open for input · Category: Big bet

Summary

Optionally let independent TAS deployments exchange behavioral patterns — never raw data — under explicit policy (island by default, share-only, receive-only, or both). Stop reinventing the same agent improvements in parallel while keeping strict data isolation; every imported pattern still lands as a reviewable PR with its origin attached.

Why

Every deployment re-solves the same problems alone. There's enormous shared learning locked up by (correct) data-isolation boundaries. The win is sharing the shape of a good solution without sharing anyone's data.

Sketch

• Patterns are abstracted, signed, and carry provenance — no customer data.
• Policy levels: island / share-only / receive-only / share + receive, set by an org admin.
• Relationships are bilateral or small-group — no public marketplace.
• Imports never auto-apply: they open a PR, scored for fit, labeled by source.
  (Provisional — the shape changes based on the discussion.)

Open questions

• What exactly is a "pattern" that's safe to share?
• How is provenance verified and revocation handled?
• Default must be island — how do we make opt-in deliberate and legible?

How to help

React 👍 if you want this, and comment with your use case, constraints, or pushback. This is a draft Tembo Agent Studio Improvement Proposal (TASIP); your input shapes it before it becomes scheduled work.