TASIP-002: Behavior variants

[ryw]

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

Summary

When two teams want an agent to behave differently — terse for support, chatty for sales — TAS proposes a variant rather than silently averaging contradictory feedback. Each variant keeps its own lineage and history so you can run team-specific behavior safely or reconcile later.

Why

Conflicting corrections are common and dangerous: merge them and you get mush that satisfies no one. The honest answer is sometimes "these are two valid behaviors," and the platform should make that explicit instead of hiding it.

Sketch

• Detect when incoming corrections diverge from the current behavior in incompatible directions.
• Offer a variant (a branch/snapshot with its own identity) instead of a single averaged change.
• Variants are first-class: runnable, comparable, auditable, and reconcilable.
  (Provisional — the shape changes based on the discussion.)

Open questions

• How is a variant scoped — per team, per label, per caller?
• When (if ever) do variants get reconciled back into one?
• How do we keep the variant count from sprawling?

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.