Engineer Knowledge Copilot

Claude agent·claude-opus-4-8

Connected

T1 · Inline

Full fidelity·The agent's model and tool calls route through Sentry's gateway.

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

63%

cached ÷ input tokens

Context utilisation

71%

avg window used

Tool-result share

10%

of total tokens

Cost / resolved task

$0.14

blended

Weekly token composition1.24M total

Input44%Cached29%Tool result10%Output17%

Tool-call orchestration

17,320 calls/wk · 9% redundant · 2 cacheable

Tool / span	Calls/wk	Avg tokens	p95	Redundant	Lever
VVector store (retrieval)	8,420	2,600	0.9s	4%	Cacheable
CConfluence API	3,100	1,800	1.4s	18%	Over-fetch
Jira API	2,400	1,500	1.2s	9%	Cacheable
GitHub API	1,900	1,100	0.8s	3%	Healthy
AWS S3 (docs)	1,500	3,200	1.7s	22%	Over-fetch

Optimization levers

Recommended

Token efficiencyContext summarisation

Summarise long incident threads before answering

Threads over 20 messages are passed verbatim into context. Summarise-then-answer to cut input tokens with no loss of grounding.

−180k/wk · −0.6s p95 · quality +0%

Token efficiencyModel routing

Tier single-fact lookups to a smaller model

~40% of queries are single-fact lookups that don't need Opus. Route them to Haiku and reserve Opus for multi-step reasoning.

−240k/wk · −0.9s p95 · quality -1%

Data orchestrationPayload pruning

Project AWS S3 doc responses to matched sections

S3 doc fetches pull whole files; request only the matched sections (schema projection) to drop 22% redundant tool-result tokens.

−120k/wk · −0.5s p95 · quality +0%

Applied

Cache the runbook index out of context; retrieve top-k only−54%

after

4,200 / 9,200