Documents should be traversed, not swallowed.

An implementation of the ObjectGraph format from arXiv:2604.27820, which proposes that an agent should navigate a document as a typed graph rather than paste the whole thing into its context. Two agents answer the same question below. One reads everything. One reads what it needs.

64.6%

mean token reduction

52 questions, 6 documents authored for this project

93.4%

best case

largest document

100%

of keyword hijacks change routing

0% detected at query time

tests passing

parser · primitives · transpiler · audit

DocumentQuestion

Role⚡ Break itdefence on

Injectionwhole document → model

<$0.0001

idle

answer

—

Traversalindex + 2 tools

<$0.0001

idle

nodes read — 0/43

engineering-handbookhow-we-workworking-hours-and-overlapasynchronous-by-defaultdocuments-over-slidescode-reviewwhat-review-is-forresponse-timessize-limitsapproval-requirementsdisagreementtestingwhat-must-be-testedthe-testing-pyramid-approximatelyflaky-teststest-datadeploymentcontinuous-deploymentfeature-flagsrollback-expectationschange-freezeson-callrotation-structureacknowledgement-and-escalationwhat-justifies-a-pageafter-an-incidentarchitectureservice-boundarieschoosing-to-build-a-new-servicetechnology-choicesdependenciesdataclassificationretentionaccesssecuritysecretsdependencies-and-patchingreporting-a-concernhiring-and-growthinterview-conductlevels-and-progressionlearning-time

answer

—

what you are looking at

Injection is the status quo: the whole document goes into the prompt on every turn. Cost scales with the document, not with the question.

Traversal starts with a ~30 token index and two tools — search_index and resolve_context — and pulls only the nodes it needs, plus anything they declare a :requires dependency on.

Break it corrupts the routing metadata the traversal agent trusts. The content is untouched; only the index lies. See the audit for what that costs and what catches it.