Type 3Hypothesis

Delegated-authority broker for agent tool calls

proposed
voting
running · 17d
measuring
concluded

Preregistration · v1 · 3 Aug 2026 · immutable after start

An extension creates a new version rather than editing the old.

Hypothesis

A broker that mints per-task, resource-scoped, twenty-minute tokens on behalf of a named user lets an agent complete the banking servicing pattern's twelve tool calls with zero standing credentials and under 80ms added latency per call.

Kill condition

Kill if any of the twelve tool calls requires a standing credential to complete by day 10, or if added p95 latency exceeds 250ms.

MethodBroker in front of a synthetic core-banking API and a document store, both with synthetic records. Agent requests authority per task; broker checks the user's delegation, mints a scoped token, logs the grant. Measure completion, latency, and the red team's attempts to escalate scope across the twelve calls.

Expected cost$500 in tokens, two people for three weeks

Expected duration3 weeks

Running notes · fed from harness sessions and by the pair

manual29 Aug 2026LF Lena Fischer
Escalation attempt 3 of 8 succeeded: a token scoped to one customer read a linked account through a join the API exposes. Scope has to be resource-set, not resource. Fix in progress.
harness21 Aug 2026TO Tom Okafor
broker + synthetic APIs up; 9 of 12 calls completing scoped; added p95 71ms
manual3 Aug 2026LF Lena Fischer
Design-review basis, not the tenant we never got (g-auth-broker-oauth-only). Predicted: completes, latency 60ms, one escalation found and closed, confidence 0.5.

Harness notes are auto-captured from Claude Code sessions: model, date, commit, session reference. Never the transcript, code or paths.

Artifacts

repoauth-broker POC + synthetic banking APIsrepo

measCompletion, latency and escalation attempts across twelve callsmeasurement