All signups since inception (2026-03-25 to 2026-10-22), excluding the 5 heaviest users by completed tasks. Based on 700 users.
84%
of signups never did anything. 586 of 700 people created an account and stopped. Only 9 of them came back after day 1. The product is not failing the people who connect an agent. It is failing to get people to that point.
Signups
700
excl. top 5 power users
Connected an agent
43
6% of signups
Created a task
106
15% of signups
Completed a task
37
5% of signups
Completed on 2+ days
8
22% of those 37
Most users stop before anything happens
Where each of the 700 signups ended up. A user is counted once, in the furthest stage they reached.
The 58 people in the second row wrote a task and nothing picked it up. They showed intent, so they are the best group to rescue first.
Show data
Conversion improved after June, but is still low
June brought 385 signups and almost none of them activated. Recent months convert better, with October the best so far.
Signups per month
Share of each month's signups that…
connected an agentcompleted a task
May to October only. March and April had 4 and 8 signups, too few to show a rate.
Show data
The first hour decides it
Of the 43 people who connected an agent, how long after signup they did it. Median: 45 minutes.
79 of the 106 people who created a task did it within an hour of signing up, and the median time to a first completed task is about 1.3 hours. That makes the first hour the window that matters most. 5 people did connect after day 7, so a late nudge still reaches some.
Show data
If the first task isn't picked up, the user is gone
What happens later depends on what happened to the first task.
Users whose first task stayed "not started"
Only 1 of 47 ever completed any task.
Reached 3+ completed tasks, by who created the first task
Both groups had a first task that was completed. Small samples (22 and 10 people), so a lead and not proof.
Of the 12 people with 3+ completed tasks, 10 had tasks created by their own agent (the agent logs its work in AgentRQ). They also chat more: 4.3 human messages on average against 1.1 for people with 1 to 2 completions.
Show data
People who succeed once rarely stay
The 37 people who completed at least one task.
How many tasks they completed
When they last completed one
Only 8 completed tasks on 2 or more different days, 3 on 3 or more, and 1 on 5 or more. 10 of the 12 people with 3+ completions have gone quiet for over 14 days. Almost nobody has push (about 5) or Slack (4) connected, so there is no way back in except email.
Show data
Claude Code users complete tasks most often
Of people who connected through each client, the share who then completed a task. "node" is a generic client and hides what the user actually runs.
Other clients (Go-http-client, Python-urllib, undici, omp) combined: 5 of 6. Permission prompts did not get in the way: 14 people saw manual prompts and 9 of them completed tasks.
Show data
What to do, in order
Rescue the 58 people with an unpicked taskEmail "your task is waiting for an agent" with the one-line connect command, and show the same message on any not-started task in the app. Highest intent, lowest effort.
Make the first hour workPut the connect command on the first screen and on the task page, not behind a setup guide. Add a guided starter task so the first run succeeds quickly. "Create first task" should not wait for an agent.
Send a signup sequence triggered by activityIf there is no agent connection after 1 hour, send setup help. Follow up at about 24 hours and day 3, and send one late nudge at day 7. The email pipeline exists today but only sends task notifications.
Teach the agent to log its own workAfter connecting, offer a snippet or skill that tells the agent to create and update tasks. Test it on a cohort before treating the pattern as proven.
Prompt after the first winWhen the first task completes, offer push or Slack, a second workspace and a scheduled task. Send a day 2 or 3 follow-up. A second workspace is a marker of engagement: 9 of 37 completers have one, against 3 of 586 users who did nothing.
Make Claude Code the default path and fix agent namingLead with its one-liner. Record the real agent name instead of the launch command, so "node" stops hiding what people use.
Fix the measurementAdd a completed-at column, an activation event, and events for setup guide opened and command copied. Count agent connections made over the ACP route too.
On the 80% goal
Messages alone will not reach 80% of signups. 84% do nothing and nearly all of them never return, so most will not read what is sent to them. A one-line connect step on the first screen is the change most likely to move the number.
I would define "engaged" as completed a task within 7 days of signup, measured for real signups, and set staged targets. For scale: lifting the connect rate from 6% to 15% (the level of the stronger recent cohorts) would add about 60 activated users per 700 signups.
Method and limits
Excluded the 5 users with the most completed tasks (88 to 3,176 each). 700 of 705 signups remain.
"Connected an agent" is the MCP connect telemetry event. People using the ACP or gateway route may not emit it, so "never connected" is probably overstated.
The schema has no completed-at column. Users were counted as completing when a task's status is "completed"; dates use the task's last update.
Many groups are small (10 to 40 people). Treat the agent-created-task and client findings as leads.
Names and emails are left out on purpose. Counts only.