SeenRelay — measuring redundant external reads across agent workers before trying to optimize them

[Vlad Belciug (ovladon)]

I’m looking for real multi-agent workloads where independent workers may be paying to read or validate the same external state.

The problem is narrower than general caching. Examples are overlapping paid web searches, scraping/browser calls, metadata/status reads, or other deterministic read-only operations performed by separate workers.

SeenRelay now has a local-only GitHub Action that can prescreen a repository without connecting to the SeenRelay service:

uses: ovladon/seenrelay-cost-prescreen@v1

If a workload survives that prescreen, the next step is shadow measurement. Every authoritative operation still runs. The measurement compares local single-flight/cache, source/provider-native controls, exact recurrence, agreement, latency, cost provenance and coordination overhead.

Only after that can a workload receive:

USE / DO NOT USE / INSUFFICIENT EVIDENCE

I’m specifically looking for real Agent Framework workloads to test this boundary. I am not assuming SeenRelay will win; proving that a native mechanism is better is equally useful.

SeenRelay:
https://github.com/ovladon/seenrelay

Cost Prescreen:
https://github.com/ovladon/seenrelay-cost-prescreen

[Andre Vianna (AndreVianna)]

I’ve been struggling with this in a coding project. The implementation moves forward, but the review and testing keep taking longer and costing more than I expected. In my case, repeated model context is part of the problem. I understand that’s different from the external reads you’re looking at here.

My question is about how fast the code changes. Several agents might review the same commit, but then one fix changes what the next agent needs to see. Have you tested that kind of workflow? Do you still get useful savings, or does keeping everything up to date eat most of the benefit?