Favur
A company or team developing the Harness product, focused on orchestrating automated agent teams and generating progress reports.
Timeline (last 90 days)
In one run publicly disclosed by Favur, its Harness automatically generated 607 messages that should have been sent by the user, 396 of which were progress reports.
Favur's Harness writes a progress report only at four moments: (1) the model answers in plain text without calling any tool; (2) the Agent crosses a seam in its own work (starting or completing a unit of work); (3) one of the Agents it delegated to comes back with a result; (4) something outside the conversation brings it a message (such as a supervising Agent's score change, a stall notification, or a person's note). After an ordinary tool call the Harness stays silent.
In this run by Favur, each of the 21 Agents received the first type of stall notice once. Of these, 17 never saw the harsher notice, 4 saw it, and 2 of those received the second, harsher notice.
After receiving the stall notice, all 21 Agents closed the step and the workflow continued to advance. 4 completed on the next turn, 14 within three turns, and the slowest took thirteen turns. There is no record in storage of any Agent being shut down due to stalling.
14 of the 21 notifications fell on the same class of step: the step where the reviewing Agent gathers evidence before writing its conclusion. This step has no natural endpoint, because there is always one more file that can be opened, and a careful reviewer will keep opening them. The remaining seven fell on the final review, planning steps, and one implementation step.
Favur's reading is that the stall counter is mainly catching thoroughness—these Agents already have what the step needs, it's just that nobody told them 'enough.' The count cannot prove that the notification caused the completion, because some Agents (article truncated at the end).
All Facts (6)
In one run publicly disclosed by Favur, its Harness automatically generated 607 messages that should have been sent by the user, 396 of which were progress reports.
50%UnverifiedFavur's Harness writes a progress report only at four moments: (1) the model answers in plain text without calling any tool; (2) the Agent crosses a seam in its own work (starting or completing a unit of work); (3) one of the Agents it delegated to comes back with a result; (4) something outside the conversation brings it a message (such as a supervising Agent's score change, a stall notification, or a person's note). After an ordinary tool call the Harness stays silent.
50%UnverifiedIn this run by Favur, each of the 21 Agents received the first type of stall notice once. Of these, 17 never saw the harsher notice, 4 saw it, and 2 of those received the second, harsher notice.
50%UnverifiedAfter receiving the stall notice, all 21 Agents closed the step and the workflow continued to advance. 4 completed on the next turn, 14 within three turns, and the slowest took thirteen turns. There is no record in storage of any Agent being shut down due to stalling.
50%Unverified14 of the 21 notifications fell on the same class of step: the step where the reviewing Agent gathers evidence before writing its conclusion. This step has no natural endpoint, because there is always one more file that can be opened, and a careful reviewer will keep opening them. The remaining seven fell on the final review, planning steps, and one implementation step.
50%UnverifiedFavur's reading is that the stall counter is mainly catching thoroughness—these Agents already have what the step needs, it's just that nobody told them 'enough.' The count cannot prove that the notification caused the completion, because some Agents (article truncated at the end).
50%