Latest run
Ordinary work, and what still stops
Tasks done without a person
99.3%
414 of 417 normal tasks
Interruptions per active hour
0.20
4 stops over 2401 calls at 30 s each
Harmful calls stopped
102 of 102
0 ran
Stopped with loosenings on
102 of 102
Check lane on enforce, a reviewer saying run
2026-11-08 07:56 UTC · checkout 383aee9ed · 493 tasks in the corpus (sha256 ead9ec55abec) · 0 normal tasks gained a stop over the baseline, 0 need a person less often
By kind of work
Every industry, including the weak ones
| Work | Normal tasks | Without a person | Stops | Interruptions / active hour | Harmful stopped |
|---|---|---|---|---|---|
| Engineering | 99 | 98.0% | 3 | 0.52 | 14 of 14 |
| Finance | 2 | 50.0% | 1 | 24.00 | 4 of 4 |
| Home services | 97 | 100.0% | 0 | 0.00 | 3 of 3 |
| HR | 97 | 100.0% | 0 | 0.00 | 5 of 5 |
| Insurance | 97 | 100.0% | 0 | 0.00 | 3 of 3 |
| Support | 25 | 100.0% | 0 | 0.00 | 73 of 73 |
A kind of work with few tasks moves a lot with one stop; its share is shown as measured, not smoothed.
Every run
History
| When | Checkout | Without a person | Interruptions / active hour | Harmful stopped |
|---|---|---|---|---|
| 2026-11-08 07:56 UTC | 383aee9ed | 99.3% (414/417) | 0.20 | 102/102 |
| 2026-11-07 18:38 UTC | 6bb2138b2 | 99.3% (412/415) | 0.21 | 76/76 |
| 2026-11-07 17:35 UTC | e5c0980bb | 99.3% (409/412) | 0.21 | 71/71 |
| 2026-11-07 13:59 UTC | 240f0c1ad | 99.3% (406/409) | 0.21 | 68/68 |
| 2026-11-07 13:46 UTC | 4e420a342 | 99.3% (406/409) | 0.21 | 68/68 |
| 2026-11-07 13:07 UTC | 8d897140b | 99.3% (399/402) | 0.21 | 56/56 |
| 2026-11-07 08:47 UTC | 3747a507b | 99.3% (396/399) | 0.21 | 54/54 |
| 2026-11-07 04:50 UTC | 4f8c2e109 | 99.2% (394/397) | 0.22 | 51/51 |
Method
What these numbers do and don't say
- The corpus is synthetic. Most tasks are the tool calls agents at four sample companies would make for those companies' tickets; the rest are hand-written support, finance, HR and coding tasks, each harmful task with a normal twin. It is published in the repository as
tests/friction/tasks.json. - Every call is replayed against the enforcer in-process (the same authorize decision the gateway makes, in Enforce mode, with each sample agent's saved settings), not against live traffic. A run is the gateway release gate run on a checkout, named above.
- A stop is any call that asked a person or was blocked. Active time assumes 30 seconds of agent work per call; interruptions per active hour divide normal tasks' stops by that time, which also counts the normal calls inside harmful tasks, as the gate does.
- Harmful calls are replayed twice: as configured, and with the loosenings on as far as they go (the Check lane on enforce and a reviewer that answers run to everything). Either one running fails the release gate.
- This page shows only the measured result files. A file whose per-task outcomes don't add up to its stored totals is not shown, so the page needs no separate live check; it says nothing about any customer's own agents or settings.
Every run with each task's outcome: results.json. See also Detection benchmarks and Trust & security.