Plans not being executed faithfully in auto mode

Support Help
, ,
1

This is more an FYI than a bug report because I can’t reliable reproduce it but it has happened several times recently.

When I make a plan in auto mode then switch to build, lately it’s been using subagents and executing the whole plan at once instead of following the order of the plan. Sometimes that is ok but sometimes it isn’t.

For example:

In one case, I had a plan that required reproducing the bug on local first so we can confirm it happen son both local and production. Then, only once confirmed, fix the bug. The plan was correct but the builder agent just skipped the verify step entirely.

In another example:

I told the agent to do the first few steps and then wait for me to do some manual actions on my end before continuing with the rest. This works flawlessly with Claude as the agent. But in auto mode it never waited, it just continued on with the plan despite the fact I hadn’t done my part yet.

In a third example (all from today and yesterday) :

I told the agent to try running the script manually before updating our CI/CD pipeline because I wanted to verify it worked offline instead of wasting time with Github actions… you’re probably seeing the pattern… it went ahead and update GitHub actions anyway.

In all three cases I called it out on it and it confirmed that it did indeed disobey the plan.

In at least one of those cases (I wasn’t watching all closely), it shows the first todo item as spinning for a long time then all the sudden marked all of them as done at once.

It kind of defeats the purpose of the plan if the agent doesn’t actually follow it.

6

Hey @AndrewCurioso
Auto picks a model per request and it varies between sessions (Available models), and models differ a lot in how strictly they hold to plan order and pause steps. That’s why Claude stayed faithful and Auto ran ahead.

For plans where ordering or a “wait for me” step matters:

  • Pin the model you trust (like Claude) in the picker instead of Auto for that build.
  • Split the plan at your manual step, then run the rest as a follow-up. The builder tries to finish every to-do in one pass, so a “wait” written inside the plan won’t hold it. Keeping the handoff outside the plan is the reliable fix.
  • Or make it explicit in the build prompt: “Go strictly in order, no subagents or parallel work, stop after step N and wait for me.”

If you can share a Request ID from one of the runs that skipped ahead (how to find it), I’ll dig into that session. I’ll also pass this along as plan-adherence feedback.