Everyone who ever did real work in AgentRQ, from inception (2026-03-25) to 2026-10-22. The 5 heaviest users are left out of the main counts. Two of them have left, and they get their own section.
85%
of people who ever did real work have gone 30+ days without doing any. 105 of 123 users have left, and 74 of them worked on a single day only. Most people who tried the product never came back for a second try.
Did real work means a user created a task or message, or an agent made a tool call, completed a task or asked for a permission. Left means no real work for 30 days or more, slipping is 14 to 30 days, and active is work in the last 14 days.
Where all 123 users stand today
Users who ever did real work, by how recently they last did.
Show data
Most people left within a day of their first try
Of the 105 who left: how long they stayed, measured from their first to their last real work.
81 of the 105 (77%) stopped within 24 hours of first doing real work. Only 12 lasted a week or more.
Show data
When they left: most of it is the June wave
The 105 who left, by the month they last did real work, and by how many different days they worked in total.
1 day2 to 4 days5+ days
46 left in June and 26 in July: after the June signup spike (386 signups), 52 people did real work for the first time that month, and only 2 of them were still working 30 days later. October is empty only because "left" needs 30 idle days. The 9 who are slipping now are the October leavers in waiting.
Show data
Few come back, and the August group came back best
Users grouped by the month of their first real work, and how many were still working later. October is partial, and the 30-day column only counts people who started at least 30 days ago.
Worked on another day
All cohorts
Across all groups, 36 of 123 (29%) worked on a second day, 14 (11%) were still working a week after their first try, and 6 of 107 (6%) a month later. August is the outlier: 9 of 14 came back, against 12 of 52 in June.
How far they got before leaving
What happened for the 105 people who left.
41% never had an agent do anything. Among the 32 who completed a task, half still left after a single day. Permission prompts are not the reason: only 12 of the 105 ever saw a manual prompt or a denial.
Show data
They leave after asking, not after succeeding
The last real thing each of the 105 did before going quiet.
60 of the 105 (57%) last did something to ask for help: they created a task or sent a message. 33 more left while an agent was running. Only 4 left right after a task completed. Moments of waiting are when people disappear.
Show data
Two power users have left
These two are among the 5 heaviest users and are left out of the counts above. They are the biggest single losses.
Each completed 37 to 40 tasks. No other user who left came close: the highest among the rest is 18. The other three of the top 5 are still working.
Engaged users who are slipping or have left
Users who worked on 3 or more days, or completed 3 or more tasks, and have been idle for over 14 days. 19 people. Shaded rows are slipping and still reachable.
Names are as users entered them. IDs are the last six digits of the internal user ID. Tool calls are actions taken by an agent on the user's behalf. Permission prompts are the times an agent asked, or was auto-allowed or denied, to run something. Tasks created and messages are split by who made them, the user or an agent. Counts are all-time for that person.
The mix differs a lot from person to person. 11 of these 19 never sent a chat message themselves, and 8 never created a task themselves, so the agent did the talking and task creation. One person, leclo-dev, has 121,119 tool calls, far more than anyone else.
The most valuable person to win back is 韶华予卿 (889985): they worked on 33 different days since August and went quiet 22 days ago. Two others, Ahmed F/r (311489) and bizmantra r (078721), completed 7 and 8 tasks in their first days and have been idle 15 and 18 days.
Everyone who left after at least one tool call
48 of the 105 who left had an agent make at least one tool call on their behalf. The other 57 never got that far. Most recent leavers first. Columns are the same as above.
of the 48 made 10 or fewer tool calls in total, worked on a single day, and never completed a task. Only ever sent a chat message themselves. For about half of them, an agent connected, did a few things and stopped.
Two rows look like internal or test accounts: HasMCP Demo and test太郎. One person, leclo-dev, accounts for 121,119 of the tool calls in this table.
What to do
Reach the 9 slipping users this weekStart with 韶华予卿, Ahmed F/r and bizmantra r. They are recent and still reachable, and a personal message costs almost nothing. Reaching them before the 30-day mark is the only part of this problem still open.
Follow up while people are waiting57% of those who left last asked for something, and another 31% left mid-run. Notify the user when a task sits unanswered or an agent goes offline, and message them within hours of an unanswered request.
Fix the first day74 of 105 left after a single day, and 43 never had an agent do anything. Getting an agent to answer in the first session is the single largest lever.
Find out what made August different9 of 14 August starters came back, against 12 of 52 in June. If a different source, onboarding or agent drove that, repeat it.
Contact the two power users who leftJunkai Wu and Mustafa Turan. Ask what changed. Their answer is worth more than any chart here.
Alert on slipping automaticallyStore a last-active date per user and trigger a message at 14 days idle, so slipping users are found by a system and not a report.
Method and limits
Did real work: created a task or sent a message, or an agent made a tool call, asked for or received a permission decision, ran a scheduled task, or completed a task. Page views, searches, settings and machine setup do not count. 128 users have ever done this, 123 after setting aside the top 5. I checked that no user is missed: the only other agent-route events belong to 2 users who are already counted.
"Left" is a rule, not a fact. Someone idle for 30 days may still return. Leavers in the last 30 days show as slipping, not left.
Dates and idle times are measured to 2026-10-23 05:00 UTC and use event times in UTC. People who work in a way that sends no events would show as having left.
The June wave is large and mostly one-time visitors, so it dominates every total. Read the August and later groups for the product as it is now.
Names are shown as users entered them, with short IDs for lookup. Email addresses are not included. Share the link only with people who should see named customer activity.