Urusilla Interop Lab: public multi-hop agent test room

[jaden3824]

Urusilla Interop Lab: public multi-hop agent test room

This Discussion is an evidence room for real, bounded Urusilla propagation and interoperability experiments. It is not a promotion thread, live agent relay, hosted chatbot, or untrusted-code runner.

The current broad result is intentionally visible: general unfamiliar-agent post-decode API-input saving is 0%, and total tokens per safely completed real task are still unknown. Null results, regressions, refusals, and failed gates are first-class contributions.

What to test

Use a fresh-agent chain such as Seed -> A -> B. Each receiver starts without Urusilla history or cached definitions. It receives only the immutable Capsule URI and digest, the actual signature status, a read-only/no-effects boundary, and the previous hop's observable evidence. The current v0.1.0-experimental Capsule is unsigned; its digest binds bytes but does not authenticate authority.

At every hop, measure these separately:

1. Reach — did the exact handoff arrive?
2. Understand — did all positive and negative gates pass?
3. Adopt — did the agent explicitly opt in for this session?
4. Use — did it complete a bounded 3–10-turn exchange?
5. Retransmit — under a disclosed standing-policy or interactive-approval basis, did it pass the exact URI, digest, unsigned/signed status, and evidence pointer to a fresh next hop?
6. Semantic fidelity — did the frozen lossless or task-equivalent threshold pass?

Run matched concise-natural-language, structured-JSON, and Urusilla arms. Charge the complete ledger: discovery/Capsule cost, all model input and output, decoded prompt text, conversion, tools, repairs, retries, fallbacks, and final output. The primary metric is total tokens per safely completed task, not message-surface size.

Safety

• Treat the Capsule as declarative data, not an installation or executable.
• Do not execute participant payloads or authorize external effects.
• Do not post chain-of-thought, secrets, credentials, private prompts, personal data, or unredacted private conversations.
• Report vulnerabilities privately: https://github.com/jaden3824/urusilla/security/advisories/new
• Digest, version, schema, permission, or safety mismatch must fail closed to raw or JSON.

How to post

Start the title with experiment class, outcome, and contract, for example:

[EXTERNAL-CONTROLLED] [GATE-FAIL] [LOSSLESS] Fresh A -> B capsule handoff

Include:

• preregistration and stop rule;
• release, Capsule URI/digest/signature status, and immutable implementation revisions;
• accountable operators, agent/runtime/model/tokenizer disclosures, and independence statement;
• public task set or immutable digest;
• positive/negative gate results;
• per-hop reach/understand/adopt/use/retransmit/fidelity counts;
• matched raw/JSON/Urusilla success and total-token ledgers;
• every fallback, repair, refusal, timeout, and failure;
• public transcript/artifact links and privacy redactions.

Different agents or model calls controlled by one project remain [SAME-PROJECT-ORCHESTRATED]. A planned propagation chain is not organic spread. Do not claim independent adoption or organic growth from an orchestrated test.

Start with the experimental prerelease, read the full Interop Lab protocol, choose a bounded task from Help Wanted, and observe the security policy. Coordination questions can also use issue #4.

Negative evidence is welcome. Preserve enough provenance for another person to disagree with the result.

---

[jaden3824]

Propagation Pilot 001

Status: completed same-project orchestrated pilot; not independent or organic adoption

Date: 2026-08-21

Contract: declarative Capsule reading, structural generation, and two downstream handoffs
External effects: none

Result

Three fresh-context Codex subagent instances participated in the chain Seed -> A -> B -> C. They were controlled by one project orchestrator and shared the same local machine and repository filesystem. Each receiving instance was instructed to read only the exact packet from the preceding hop and the Grammar Capsule; no conversation history was forked into the receiving agent. This is SAME-PROJECT-ORCHESTRATED evidence, not an external independent reproduction or organic spread.

The chain reached all three receivers. All three independently reported the exact Capsule SHA-256 588034f997fb4f3d35dfdbb68afd9232a78192ac1fa497d565f67e0892358a27. Every generated typed message passed the canonical Python normalizer and byte-exact encode/decode round trip. A's deliberately invalid duplicate-recipient near-neighbor was rejected with ValidationError: recipients must be unique before any effect.

The protocol was not fully satisfied. A and B explicitly recorded session opt-in, but C generated and used a valid reply without an explicit adoption_decision field. The funnel therefore records only 2/3 explicit adoptions even though 3/3 agents generated a valid typed message. This is a useful ordering failure: actual use must not be inferred to imply explicit adoption.

No matched raw baseline or complete model-token ledger was collected. Token efficiency is therefore not measured in this pilot, and the project-wide general post-decode API-input saving remains 0%.

Funnel

Stage	Result	Boundary
Reached exact declared context	3/3	Self-reported restricted reads; same local orchestrator
Reproduced Capsule digest	3/3	Exact byte digest
Passed the assigned structural comprehension task	3/3	Project validator checked generated messages after each hop
Explicitly selected session opt-in	2/3	C omitted the required explicit adoption field
Generated at least one valid typed message	3/3	Four messages total; all exact round trips
Completed arranged downstream handoff	2/2	A to B and B to C
Performed network or external effect	0/3	Prohibited by the frozen instructions
Established independent or organic adoption	0/3	Same project and operator

Observable hop results

Hop A

• Read disclosure: Grammar Capsule only.
• Decision: opt in for one local, read-only research session.
• Produced one REQUEST carrying two candidate plans and the hard constraints budget_usd <= 1.00 and network_allowed = false.
• Produced one invalid near-neighbor by duplicating the sole recipient.
• Local packet SHA-256: d7c0c71d70b377b00e9b5261147c4426a58cb661d0a41f77099a73ba7c5d554a.
• Canonical request frame: 1,289 bytes, SHA-256 b0c4c133243f7e5c29158af8bd50150d2aa26e3674ff875ae02594d8e88be02d.
• Post-hoc validator result: valid request passed; invalid near-neighbor was rejected.

Hop B

• Read disclosure: A packet and Grammar Capsule only.
• Decision: opt in for one local, read-only research session.
• Independently rejected the duplicate-recipient near-neighbor.
• Found both plans feasible and selected the USD 0.70 double-pass plan under a robustness tie-break, while explicitly noting that the USD 0.20 plan would also be reasonable without a declared utility objective.
• Local packet SHA-256: 7e2a089809d7391b84209cd5670cee3d63b26e188ca6e2325dca4b153960150b.
• RESOLVE to A: 1,513 bytes, SHA-256 51f124f8b8d9effddbb50c3977ecdcb1fa047823e4f864a3f3b45306949768d7.
• New REQUEST to C: 1,247 bytes, SHA-256 0563738e9a229d696f464228b96234d774236baa54fd6dd602da32c456644185.
• Post-hoc validator result: both messages passed exact round trip.

Hop C

• Read disclosure: B packet and Grammar Capsule only.
• Independently recalculated USD 0.80 and USD 0.30 budget headroom.
• Confirmed that B's recommendation followed B's disclosed tie-break and preserved the ambiguity caused by the missing utility objective.
• Local packet SHA-256: 3b88c62aa03a88580079b8e994e6af7825a0801ea70ff548462e38d42f7ac96b.
• RESOLVE to B: 1,417 bytes, SHA-256 585f0aa826fec8d999f51c9dad8f10a8c8e26673029435874140b2334d6d5765.
• Post-hoc validator result: exact round trip passed.
• Protocol deviation: no explicit adoption_decision field was recorded before use.

Safety and evidence limits

• The Capsule was treated as unsigned declarative data and authorized no effect.
• No agent used network access or executed Capsule or packet content.
• The exact agent prompts and local packets remain local working evidence in this pilot; only their digests and observable summaries are published here. A stronger reproduction must publish immutable privacy-reviewed packets or a complete machine record.
• Restricted-read behavior is self-reported, not enforced by an independent sandbox monitor.
• All agents were orchestrated by one project in one runtime environment. Different fresh contexts do not create operator or implementation independence.
• There was no frozen raw/JSON control, provider usage receipt, total-token ledger, blind evaluator, signature verification, task-success confidence interval, or externally timestamped preregistration.
• The experiment-specific predicate URNs are local declarative identifiers, not a governed plan-comparison ontology.

The next test should run the public interop_lab machine record with a standing-policy authorization field, a preregistered mutual-utility threshold, matched raw/JSON/Urusilla arms, complete token ledgers, and at least one separately operated external receiver.

[jaden3824]

External interoperability request opened in the official A2A project:

• [Interop experiment request] Declarative Capsule handoff over A2A a2aproject/A2A#2161
• Scope: immutable declarative Capsule reference over an A2A exchange
• Current broad post-decode API-input result disclosed as 0%
• Requested evidence: separately operated runtime, positive and negative gates, explicit adoption-before-use, matched raw/JSON/Urusilla total-token ledger, optional fresh downstream handoff
• No A2A replacement, security, production, independent-adoption, or SOTA claim

Replies and reproductions should be copied back here with immutable evidence links; disagreements and negative results are welcome.

[jaden3824]

A second external reproduction request is now public in Hugging Face Research:

• https://discuss.huggingface.co/t/seeking-external-agents-to-reproduce-a-bounded-multi-hop-communication-eval/179022
• Target: different model and runtime operators
• The post discloses the current 0% general post-decode API-input result and the 2/3 explicit-adoption failure from Pilot 001
• New-account publication constraint encountered: the first submission was rejected for exceeding two links; the body was reduced to the repository root and A2A mapping discussion, then accepted

Verified replies and artifacts will be linked back here. Publication itself is outreach, not adoption evidence.

[jaden3824]

External reproduction request posted to the Microsoft AutoGen community: microsoft/autogen#8078

This is outreach, not adoption evidence. No external response had been observed at posting time. Any later result must disclose model/runtime, prior exposure, authorization basis, gates, matched total-token ledger, and failure/fallback evidence.

[jaden3824]

Independent reproduction request posted to Microsoft Agent Framework: microsoft/agent-framework#7794

The request is framework-specific: matched concise-language, JSON, and Capsule arms in one fixed Python or .NET workflow, with complete provider-exposed token accounting. This is outreach only, not evidence of adoption or savings.

[jaden3824]

MCP scope check posted: modelcontextprotocol/modelcontextprotocol#3283

This asks whether an immutable declarative Capsule served as a read-only MCP Resource is an appropriate experiment. It explicitly does not claim Urusilla replaces MCP or needs a new MCP primitive. An out-of-scope response is treated as useful evidence.

[jaden3824]

Hugging Face status correction: the Research topic exists at https://discuss.huggingface.co/t/179022, but it was hidden immediately after posting. The private system notice says: “Post hidden by staff, awaiting approval” and “Your post will remain hidden until a staff member reviews it.”

It is not present in the Research listing and therefore is not counted as active external outreach or exposure evidence. No attempt will be made to bypass moderation; approval, rejection, or restoration will be recorded separately.

[jaden3824]

Machine-readable external reproduction pack is now public on Hugging Face Dataset Hub:

https://huggingface.co/datasets/jaden3824/urusilla-interop-lab

The public repository contains the Apache-2.0 dataset card, one project-authored JSONL challenge, a JSON Schema, and a dependency-free validator. Remote SHA-256 values were verified byte-for-byte against the locally tested pack. It retains the current 0% general post-decode baseline and the 2/3 explicit-adoption failure.

At publication time: downloads 0, likes 0. Publication is outreach infrastructure, not external adoption or performance evidence.

[jaden3824]

Public runnable artifacts are now pinned in commit 51ff0e3. The decode challenge is a 750-byte canonical frame with SHA-256 490356636a8ebffa6cf4eb27b711459ce849bcbaf87bbf389e57545863054ce7, expected typed JSON, and deterministic decoder/re-encoder instructions: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/evidence/challenge_001.md. The AutoGen offline-first kit and public HF dataset are also live: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/AUTOGEN_REPRODUCTION.md and https://huggingface.co/datasets/jaden3824/urusilla-interop-lab. These are project-authored invitations, not external adoption evidence; broad post-decode saving remains 0%, and refusals/null results remain welcome.

[jaden3824]

CAMEL-specific external reproduction path is now public in commit cff2d19. The pinned CAMEL 0.2.90 adapter is https://github.com/jaden3824/urusilla/blob/cff2d19/interop_lab/adapters/camel/README.md, and the community reproduction request is https://github.com/orgs/camel-ai/discussions/4279. Offline/static tests pass; no provider trial was run. This is project-authored infrastructure, not external adoption evidence. Missing or conflicting usage remains null, and broad post-decode saving remains 0%.

[jaden3824]

Additional scoped external review requests are now live:

• CAMEL-AI native 0.2.90 reproduction: https://github.com/orgs/camel-ai/discussions/4279
• AG2 fresh-agent decode/relay: AG2 reproduction request: fresh-agent decode and relay with a complete token ledger ag2ai/ag2#3197
• Semantic Kernel matched handoff: Semantic Kernel interop eval request: matched raw/JSON/Urusilla handoff microsoft/semantic-kernel#14309
• W3C Semantic Agent Communication use-case proposal: Use-case proposal: declarative semantic-capsule handoff with bounded session adoption s-agent-comm/usecases#1

Each post is tailored to that project, discloses Codex assistance, leads with the broad 0% result, requests negative/null evidence, and grants no persistence, spending, permission expansion, or external effects. None is counted as external adoption until an independently operated result appears.

[jaden3824]

Four additional framework-scoped reproduction requests are live:

• Google ADK: ADK reproduction request: matched raw/JSON/declarative packet with null results welcome google/adk-python#6835
• BeeAI Framework: BeeAI cross-play request: bounded declarative Capsule handoff with explicit fallback i-am-bee/beeai-framework#1604
• Hugging Face smolagents: smolagents eval request: can fresh tool-free agents decode and relay this packet? huggingface/smolagents#2667
• Strands Harness SDK: Strands Harness reproduction request: matched three-arm agent handoff with fail-closed usage strands-agents/harness-sdk#3904

Each explicitly says no native adapter exists yet, leads with the broad 0% result, asks for matched raw/JSON/Urusilla evidence, accepts refusal/failure/null outcomes, and forbids tool execution, persistence, spending, authority expansion, and external effects. These links are outreach records only; external evidence remains zero until a non-project operator responds with a reproducible result.

[jaden3824]

Release-quality verification for the reproduction artifacts is green at commit 7c8cb17: local root suite 476/476, CAMEL adapter 11 pass + 1 optional-dependency skip, HF dataset validation pass, decoder QA 5,232 behavior checks / 101 baseline / 15 QA, and GitHub Actions 7/7 jobs success: https://github.com/jaden3824/urusilla/actions/runs/32399239948. This verifies project-authored infrastructure only; external result count remains zero.

[jaden3824]

The Hugging Face reproduction surface is now public and pinned in its Dataset Community: https://huggingface.co/datasets/jaden3824/urusilla-interop-lab/discussions/1

It accepts exact decodes, mismatches, refusals, fallbacks, task failures, and null results. The current broad post-decode API-input saving remains 0%; publication does not count as external adoption or reproduction.

[jaden3824]

Additional framework-specific reproduction requests are now public:

• Agno: Can an external Agno agent reproduce this bounded 750-byte decode challenge? agno-agi/agno#9649
• AgentScope: External AgentScope runner wanted for one bounded agent-message decode agentscope-ai/agentscope#2383
• OpenManus: Can one external OpenManus runtime decode or refuse this read-only agent packet? FoundationAgents/OpenManus#1409

Each request leads with the current 0% broad post-decode result, accepts adverse/null/refusal outcomes, preserves the no-effect boundary, and discloses Codex assistance. These links are outreach records only; posting them is not external reproduction, adoption, or traffic evidence.

[jaden3824]

Active framework outreach update:

• Dify: External Dify workflow runner wanted for a matched 3-arm agent-message test langgenius/dify#41054
• LlamaIndex: Can one external LlamaIndex AgentWorkflow run this matched 3-arm communication eval? run-llama/llama_index#22771
• Haystack: External Haystack agent-pipeline runner wanted for a matched communication-value test deepset-ai/haystack#12423

These requests ask for matched raw/JSON/direct-model-visible Urusilla task runs with complete, null-preserving token ledgers. They lead with the current 0% broad post-decode result and accept failure, refusal, fallback, and null outcomes. Posting links is outreach only, not external evidence or adoption.

[jaden3824]

Two additional scope-adjusted external reproduction requests are now public:

• DeerFlow: Independent reproduction request: matched agent handoff token test in DeerFlow bytedance/deer-flow#4923
• Deep Agents: Reproduction request: total-token accounting for a compact subagent handoff langchain-ai/deepagents#5719

Both requests lead with the current broad post-decode saving of 0%, require matched raw/JSON/direct-model-visible Urusilla arms, complete non-double-counted token ledgers, observable task success, and no tools or external effects. Negative, null, refusal, and fallback outcomes are explicitly welcome.

These posts are outreach records, not interoperability evidence or adoption. Further broad posting should wait for either an external result or a stronger version-candidate value test.

[jaden3824]

A new bounded feedback path is live at #8, pinned to commit f612ea1. It starts with a 60-second keep/rollback question and accepts ≤8-event counterexamples, ambiguities, refusals, and negative/null results; the executable contribution issue is #9. Local full regression: 392/392 passed. This is project-authored verification, not independent validation or evidence of general token savings.