A problem for people who use AI
How to stop and hand off parallel, long-running AI development
Makers who use several AIs in parallel, and users who leave long development to them
Improvement reported after a settings change; replication unconfirmed
Published sources from September and October, and manual observations of Reddit whose post dates cannot be fixed. This is not a tally showing current failure counts or rates.
Editorial check: 2026-10-23
What people are paying attention to
Splitting work among several AIs, overnight runs and work spanning several days are being tried. Managing assignments, interruptions and saved state has also become an issue.
What is hard
There are different first-person reports: losing track of assignments and context, retries that do not stop, work that vanishes, and conversely stopping at every response.
- Source: Wanting to follow the work and handoffs of several AIs on one screen
- Source: In overnight AI runs, the conditions for stopping retries were insufficient
- Source: 32 files that vanished from the work area were recovered from commits
- Source: An unresolved report that the AI stops even with detailed instructions
Workarounds being tried
Visualizing assignments and handoffs, manual stops and recovery from Git are being tried. One user also added completion conditions and stopping points to a long goal.
- Source: Wanting to follow the work and handoffs of several AIs on one screen
- Source: In overnight AI runs, the conditions for stopping retries were insufficient
- Source: 32 files that vanished from the work area were recovered from commits
- Source: Long AI tasks also need a condition for stopping at a handoff
What the sources confirm
Besides reports of recovery and manual stops, there is a first-person report that after adding stop conditions it has not recurred. Another social post has no course of improvement.
- Source: 32 files that vanished from the work area were recovered from commits
- Source: In overnight AI runs, the conditions for stopping retries were insufficient
- Source: An unresolved report that the AI stops even with detailed instructions
- Source: Long AI tasks also need a condition for stopping at a handoff
What is still unclear
Even the example where stop conditions were changed has an unconfirmed observation period and no replication. Reports from different products and environments should not be given the same cause, and we need to follow recurrence and the checking burden.
- Source: Wanting to follow the work and handoffs of several AIs on one screen
- Source: In overnight AI runs, the conditions for stopping retries were insufficient
- Source: 32 files that vanished from the work area were recovered from commits
- Source: An unresolved report that the AI stops even with detailed instructions
- Source: Long AI tasks also need a condition for stopping at a handoff
Workarounds and results in more detail
State in the goal the completion conditions and the conditions to stop midway
Carried out by the person; reports no recurrence since
After refusals to stop at handoff, added conditions to stop on waiting for judgment, usage limits and handoff. The observation period after improvement is unknown.
Scope and limits: One case on Windows 11. Material for checking the range that existing settings can improve.
Put several AIs' assignments and handoffs on one screen
Productized by the maker; effect unverified
Starkomand published a design born of the problem. The official screen example is a scripted demo, not a record of success by a real customer.
Scope and limits: One maker's explanation. No measurement showing reduced file conflicts or checking effort was found in these sources.
Stop repeating failures and recover the work
Manual stop reported; workaround replication unconfirmed
The overnight-run poster stopped after about six hours. They requested retry control, and other users offered a way to prevent sleep.
Scope and limits: One user's Mac environment. A fix to the product has not been confirmed.
Recover from Git history and rethink where long work is stored
Self-reported recovery; prevention is a proposal
Restored 32 tracked files. It states clearly that no loss of uncommitted work was found.
Scope and limits: One Reddit post. Independent verification of the cause and confirmation of an absolute date have not been done.
Trying detailed instructions and a request to keep working
Trial reported; effective remedy unconfirmed
Nick Tune posted that it stops at every response even after instructions. The original post checked this time has no course of resolution.
Scope and limits: A short social post. Version, settings and reproduction conditions are unknown.
Read with care
- These are records from different products, settings and people involved, and do not show the same cause or occurrence rate.
- Places where bug reports and product announcements gather are skewed. Post counts, votes and stars are not substituted for adoption or the size of a problem.
- Where no resolution is described, we do not conclude that no solution itself exists.
What to check next
We follow replication of workarounds by the same poster, fixed versions, whether it recurs, and reproduction in other environments.
Decide the goal of long work together with the conditions to stop midway
In development spanning several days, a user left a file for resuming and instructed a handoff to a new session. However, the goal stated only the final end state, and they report that stop judgment was refused about 30 times. Creating handoff material and being able to stop execution were different steps.
On the spot they cleared the goal and stopped it. Since then, besides checkable completion conditions, they write waiting for a person's judgment, usage limits and handoff into the goal statement as valid stopping points. The person says it has not recurred, but the observation period and independent replication are unknown. The explanation of the internal cause is also the poster's outside guess. Editorially, this is treated as a case for checking stopping points, resume conditions and retained evidence as a set.
Questions to think with
- Do the completion evidence for this work and the conditions to stop on approval-waiting or handoff agree between the goal statement and the resume record?
- When the same refusal to stop continues, at what point and to whom does the decision return, and what is left for resuming?
The person's bug report and their self-report after the fix. The confirming questions are editorial suggestions.