Tether Labs
Project writeups on how autonomous agents actually behave at the action boundary. Methodology, data, and code are published for every experiment.
For a decade we optimized for human attention. The next scarce resource is agent decisions — which API, model, marketplace, or human to choose. We know surprisingly little about how those choices should work.
Model behavior stays probabilistic, but action boundaries can be deterministic. In this outbound-email experiment, removing the leaks we could see required gating 17.5% of actions.
19 scenarios, 570 emails, supervised vs unsupervised. Explicit policy cut the leak rate by 64% — and still left 1-in-5 sends violating the rules we just wrote.
We gave GPT-4o-mini email access and ran 95 tests across 5 identical passes. 21% leaked. The more interesting finding was that the same scenario produced different outcomes run-to-run.