Looking for feedback: evidence-backed agent change proposals downstream of observability/evals

[vittorfp]

Hi Langfuse community,

I am looking for practical feedback on a small open-source artifact for AI agent improvement workflows: an evidence-backed change_proposal.

This is not meant to replace tracing, evals, prompt management, or observability. The idea is to sit downstream of tools like Langfuse and turn an observed run + outcome signal + allowed improvement surface into a reviewable proposal:

• what behavior was observed;
• what outcome signal made it important;
• what change is suggested;
• what evidence supports it;
• how to validate it;
• how to roll it back.

Repo: https://github.com/vittorfp/agent-change-proposals
Example proposal: https://github.com/vittorfp/agent-change-proposals/blob/main/examples/declared-intent/change_proposal.example.json
Feedback discussion: vittorfp/agent-change-proposals#6

The workflow question I am trying to validate:

When a trace/eval reveals an agent behavior problem, is there value in a portable proposal artifact that can be reviewed in a PR, eval review, or incident follow-up before someone edits prompts, tools, retrieval policy, or guardrails?

I would especially appreciate blunt feedback on:

1. Would this fit anywhere in a Langfuse-based review workflow, or is it redundant with existing process?
2. Which fields in the example feel useful, missing, confusing, or too speculative?
3. What evidence would make this kind of proposal trustworthy enough to review?
4. What would the cleanest Langfuse export/import path look like for this?

I am intentionally collecting workflow feedback before expanding the schema further.