Trust & Governance Plane

OpenAI agent·gpt-4o

Connected

T3 · Admin API

Aggregate·Sentry pulls usage and audit records from the vendor's admin API.

Span-level analysis of every conversation this agent runs – the token economics behind each answer and the tool-call orchestration underneath it. Levers span model routing, semantic model routing, prompt caching, context pruning, schema projection, request batching, retrieval (RAG) tuning, speculative decoding and token budgeting – cutting cost and latency without moving quality.

Token economics

Cache hit rate

58%

cached ÷ input tokens

Context utilisation

52%

avg window used

Tool-result share

of total tokens

Cost / resolved task

$0.06

blended

Weekly token composition540k total

Input48%Cached26%Tool result9%Output17%

Tool-call orchestration

28,100 calls/wk · 12% redundant · 2 cacheable

Tool / span	Calls/wk	Avg tokens	p95	Redundant	Lever
AAudit log (write)	12,800	600	0.3s	2%	Healthy
SServiceNow approvals	4,200	900	1.2s	14%	Parallelizable
SSplunk query	2,600	2,100	2.4s	27%	Cacheable
OOPA policy	3,400	700	0.5s	4%	Healthy
OOkta lookup	5,100	500	0.6s	33%	Cacheable

Optimization levers

Recommended

Token efficiencyPrompt templating

Pin the audit-diff prompt as a cached template

The audit format is re-described on every run. Move it to a cached system template so it bills as cached input, not fresh tokens.

−60k/wk · −0.2s p95 · quality +0%

Data orchestrationRequest batching

Batch low-risk approvals

Low-risk approval checks fire one call per action. Batch them into a single call to cut request overhead and tokens.

−90k/wk · −0.9s p95 · quality +0%

Data orchestrationMemoised lookups

Memoise Okta identity lookups per session

Identity lookups repeat within a session; memoise per principal to remove 33% redundant calls to the Okta service.

−70k/wk · −0.5s p95 · quality +0%

Applied

Switch routine audit diffs to gpt-4o-mini; keep 4o for exceptions−51%

after

3,400 / 6,900