Can one external LlamaIndex AgentWorkflow run this matched 3-arm communication eval?

[jaden3824]

I am seeking one LlamaIndex operator outside the Urusilla project to run a bounded AgentWorkflow comparison. This is an evaluation request, not evidence of adoption or superior compression.

The current broad result is unfavorable: demonstrated post-decode API-input saving for general unfamiliar-agent dialogue is 0%. The next useful question is whether direct task-aware consumption can preserve success while reducing complete model-visible tokens.

Evaluation surface

• One-record matched raw/JSON/Urusilla task and pinned result thread: https://huggingface.co/datasets/jaden3824/urusilla-interop-lab/discussions/1
• Frozen 750-byte decode fixture: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/evidence/challenge_001.md
• Structured result issue: Good first evidence: decode challenge_001 in a fresh external runtime jaden3824/urusilla#7

Please keep the LlamaIndex model, AgentWorkflow configuration, sampling, task facts, tool policy, and success rubric identical across concise raw text, ordinary descriptive JSON, and direct model-visible Urusilla arms. Randomize or counterbalance order. Do not decode the Urusilla arm into expanded prose before the model receives it.

Report observable task success and every provider-exposed input/output, induction, repair/retry, tool, hidden-if-reported, unclassified, and total token category. Unknown fields stay null. A mismatch, refusal, fallback, task failure, or null saving is a valid result.

The packet grants no tools, persistence, cross-session memory, spending, permission expansion, network action, or external effect. Publication remains the operators separately authorized action.

A single run is bounded external evidence only; it cannot establish independent implementation, organic spread, security, general adoption, or state-of-the-art performance.

Disclosure: Codex agents assisted with the evaluation pack and this post; I reviewed and submitted them.