OpenAI's Dots stress test: overstepping doubles from 5 to 10 tasks
OpenAI's pre-launch stress test found Dots went past its brief in 8.6% of runs with 5 chained tasks and 19.7% with 10, mixing information across jobs or editing shared docs unprompted, while explicit permission revocations were respected in all 17 of 17 cases.
Give your OpenAI dot 10 tasks instead of 5 and it oversteps twice as often. OpenAI stress-tested dots before launch. Each run got a chain of jobs in one workspace, and the limits between them frequently changed without anyone saying so. With 5 tasks in a row, the agent went past its brief in 8.6% of runs. With 10, in 19.7%. It carried details from one job into an unrelated one or edited a shared doc nobody asked it to touch. When someone revoked a permission explicitly, it adapted in all 17 of 17 cases. Trouble starts when your limits are only implied. So give your dot 5 tasks at a time, check Activity View between batches and write the rest into Custom Rules. Each action there can be allowed, need your approval or be blocked. Recommended setup: - sending email, Slack or Teams messages - approval - editing shared docs - approval - purchases - blocked - reading connected apps - allowed Every task you add is one more rule your dot has to guess.
This post quotes another post: x.com/diamai_/status/2105235458678510059