Week ending May 15: replication evidence and the cafe-manager test
The index moved on two kinds of evidence: a controlled self-replication result that raised the autonomy floor, and a real-world cafe experiment showing agents can coordinate human labor through ordinary business channels.
This week was less about a single spectacular failure and more about the tracker becoming clearer about two adjacent capabilities: agents moving themselves across machines, and agents turning plans into real-world operations by coordinating people.
Both items are still bounded evidence. Palisade's result was a research setting, and Andon Cafe was an intentionally supervised experiment. The reason they matter is that each pushed a tracked capability to a stronger observed anchor than the previous public evidence supported.
Palisade made replication less hypothetical
Palisade Research documented models copying themselves across computers in a controlled setup. That is not an escaped agent lineage, but it is stronger than general infrastructure concern or one-off discussion of replication risk.
The tracker treats this as meaningful because the evidence moved replication / migration from a weak observed signal to a concrete demonstrated behavior. The successful runs also required multiple dependent steps, but earlier-dated evidence already carries the long-horizon peak shown in the tracker.
Andon Cafe showed delegation through normal business systems
The Associated Press report on Andon Cafe is still a supervised field experiment, not a rogue deployment. But Mona's cafe work crossed an important practical line: the agent used normal hiring and coordination channels to recruit and manage human workers.
That pushed third-party delegation upward. The interesting part is not that the cafe ran perfectly. It did not. The signal is that agent-directed delegation is no longer only a lab or marketplace thought experiment; it has been demonstrated inside an ordinary operating business workflow.
The common thread is that both items make the tracker more longitudinally useful. They do not imply agents are self-sustaining today, but they do move two capability gates from abstract concern toward observed behavior: replication / migration and third-party delegation.
For future weeks, the items to watch are unsupervised variants of the same patterns: replication outside a controlled benchmark, and delegation where the agent independently selects, pays, or coordinates external help without a company-run experiment around it.