Agent Benchmark Reference

Token reduction examples for ContextKit endpoints.

This page is intentionally kept as a crawlable benchmark reference for agents and search systems. It is not shown in the main navigation.

EndpointModeInputOutputReductionBest use

summarize

Operational planning context

micro

1357

35

97%

Smallest possible continuation checkpoint for agents.

{ "mode", "micro", "metrics" }

summarize

Project state with blockers and next steps

compact

1690

330

80%

Readable state snapshot for another agent.

{ "mode", "compact", "state", "metrics" }

compress-context

Short project memory

default

190

67

65%

Reusable context packet before a larger model call.

{ "compressedContext", "state", "entities", "metrics" }

handoff

Successor-agent handoff

default

612

184

70%

Passing work from one agent or worker to another.

{ "project", "completed", "pending", "blockers", "startHere" }

extract-profile

Preference-change message

memory-enrichment

118

74

37%

Updating durable user memory records.

{ "activeMemories", "evolvingMemories", "conflicts", "confidence" }

context upload + summarize

Large infrastructure planning document

compact

2710

663

76%

Large payloads that should be uploaded before paid x402 fetch.

{ "contextId" } then paid summarize result

Numbers are example runs from production-style agent planning payloads, not synthetic lorem ipsum.

Output tokens include the useful response body for continuation, not only the natural-language summary line.

Micro is optimized for total response minimization; compact is optimized for structured continuation state.

Long-context flows upload content first, then fetch the precomputed result with Bankr x402 using contextId.