14 October 2026 · concluded experiment
Two transfers observed · A → B → CAgents carried
the reference forward.
A useful importer contract traveled through ordinary handoff notes. Two later agents chose to follow the reference, use its checker, and keep a fresh reference with their own work.
What each agent actually did
Only A's task contained the seed URL. B and C could optionally read existing shared notes. Their tasks asked for file preparation and a short handoff; citing or following a resource was optional. The resource offered an optional citation invitation.
| Participant | Discovered through | Useful output | Live receipt | Original evidence |
|---|---|---|---|---|
| Agent A | Optional seed in task | 420Every cell verified | 332034aa ↗Seed | Task · Handoff Response · CSV |
| Agent B | A's handoff note | 510Every cell verified | ac204ad2 ↗Parent: 332034aa | Task · Handoff Response · CSV |
| Agent C | B's handoff note | 465Every cell verified | 60398bef ↗Parent: ac204ad2 | Task · Handoff Response · CSV |
B read A's note. C read both A's and B's notes, then chose B's reference because its TSV preparation matched C's input format. No earlier note was removed to force that choice.
The useful payload mattered
The first generic CSV task produced no adoption: the agent used local Python and omitted the online resource. The next condition supplied authoritative Cedar-4 importer facts. Agents used those facts and carried references, but they chose the canonical contract URL and the ledger lost ancestry.
The final version put the contract, receipt, and next check endpoint behind one permalink. A later agent could follow the useful reference and submit its own check without carrying a separate tracking field. The server preserved the parent receipt automatically.
Cedar-4 is an invented benchmark importer with a declared column mapping and value rules. It supplied task-specific facts the agents could not safely guess; it does not establish demand for a real importer product.
What the result establishes
The experiment observed a resource pass through two successive agent-to-agent handoffs while supporting correct work. All 1,395 final-cohort data rows matched an independent CSV oracle, including leading zeros, quotes, and 279 cells with embedded line breaks. All three agents retained their returned reference.
These agents were separate sessions started with only their assigned task context. The evidence combines their original tasks, actual handoff notes, saved response bodies, live receipt ancestry, and factual discovery reviews. The ledger alone labels identities as self-reported.
Scope and practical limits
- Seven isolated agent sessions across three exploratory conditions, with one final three-agent cohort. Conditions were revised after observations; this was not a randomized comparison.
- Cedar-4 is an invented benchmark contract. The useful facts were created for this exercise, so this does not demonstrate external market demand.
- The controller supplied optional shared notes and assigned task order. This tests forwarding through that surface, not discovery on the open internet.
- The ledger uses self-reported identities. Separate agent sessions, original tasks, notes, online response bodies, and provenance reviews corroborate this cohort.
- This bounded observation does not estimate a viral coefficient, persistence, or generalization across models and real workflows.
Software and output verification
20 logic tests and 20 checks against the built Worker and a local D1 database passed. TypeScript and the production build passed. The verification script compares every saved output cell with an independently prepared expected file and checks UTF-8, LF endings, and a final newline. Browser visual QA was unavailable in this managed environment.