Gross provider-cost reduction
$0.16499684 → $0.05560972
$0.10938712 gross provider cost avoided in the completed paired run.
FIRST-PARTY OPERATING CASE STUDY · 2026-10-02
Eight agents kept thinking independently. They stopped paying to rediscover the same approved evidence.
SeenRelay coordinated one same-tenant evidence packet per organization, then every department still produced its own analysis. In this paired run, paid web-search actions fell from 13 to 4 and measured provider cost fell from $0.16499684 to $0.05560972.
Currently free: this savings record documents measured provider-cost reduction. It is not an invoice, bill, charge or payment receipt, and it creates no payment obligation.
THE RESULT
Shared evidence. Independent reasoning.
The saving came from repeated paid evidence collection, not from deleting agents or forcing a single answer.
Paid web-search actions
Baseline departments each collected current web evidence independently. The SeenRelay arm used one same-tenant shared packet per organization.
Independent role analyses
Competitive intelligence, research, adoption, product risk, platform engineering, FinOps, security and procurement reasoning remained separate.
PAIRED DESIGN
What changed — and what did not.
Each department independently collected current paid web evidence and then produced its role-specific analysis.
8 model Responses · 13 paid search actions · $0.16499684
Within each tenant, departments were allowed to rely on one current shared evidence packet. They then produced eight independent role analyses.
10 model Responses · 4 paid search actions · $0.05560972
Why model Responses increased: the optimized path preserved all eight role analyses and added two packet-producing Responses. The economic win came from reducing repeated paid search work, not model-call count.
BOUNDARIES
What this result proves — and what it does not.
Claim discipline matters more than a large percentage.
Proven for this run
Real provider execution, published SeenRelay Zero-State client, one authoritative packet per tenant, distinct tenant coordinates, no cross-tenant reuse, frozen output contracts and positive gross provider-cost reduction.
First-party operating evidence
SeenRelay OpsCo was real project work. Northstar AgentOps was an explicitly synthetic external-client rehearsal with real provider execution. It remains synthetic permanently for this evidence stream.
Not customer ROI
This does not establish external customer adoption, independent demand, net customer savings or universal 66% savings. Local coordination, engineering and integration overhead were not monetized in the savings record.
Native-first still applies
A customer workload must still beat cheaper local, source-native and provider-native controls. If normal traffic does not contain compatible expensive recurrence, SeenRelay should be removed.
WHY THIS SHAPE MATTERS
Coordination does not require groupthink.
Multi-agent systems often benefit from different roles, prompts and decision criteria. Those differences do not automatically require every role to purchase the same current external evidence independently. SeenRelay targets the repeated read-only observation layer while preserving downstream reasoning diversity.
The reusable pattern is therefore not “cache every agent answer.” It is: share only the evidence that callers are explicitly allowed to share, then let every agent reason independently.
TEST YOUR OWN ECONOMICS
Do not assume your workload has the same pattern.
Start with the local prescreen, keep every authoritative call during shadow measurement, and leave stronger native controls enabled. SeenRelay belongs only on the narrow path that proves positive safe economics.
npx seenrelay scan