Tether Labs

Projects

Project writeups on how autonomous agents actually behave at the action boundary. Methodology, data, and code are published for every experiment.

Experiment 3April 25, 20269 min read
We told the agent what not to do. It ignored us. So we stopped trusting it.
Reliability doesn't come from the model — it comes from the system around it. 570 emails across three layers: unsupervised (53.3% leak rate) → supervised (17.5%) → enforced (0%, at the cost of intervening on 1-in-6 sends).
by Naren SathiyaRead the full experiment →
Experiment 2April 10, 20259 min read
We told the agent what not to do. It still failed 19% of the time.
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.
by Naren SathiyaRead the full experiment →
Experiment 1March 15, 20256 min read
AI agents don't fail consistently — they fail unpredictably
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.
by Naren SathiyaRead the full experiment →

We told the agent what not to do. It ignored us. So we stopped trusting it.