I use Cursor Agent with always-on workspace rules and custom skill files. My rules say zero autonomy: the agent cannot decide scope, mark steps complete, or move forward without my explicit approval.
I also built a custom slash command for planning new work. It is not a Cursor default. It is my process. Planning only, no code. The steps are:
- Copy a plan template into
docs/working-on/ - Run a scope pass and paste the result into the plan
- Stop and wait for me to confirm scope
- Only after I confirm: run more research, write the checklist, run plan verification
- Only after I approve the checklist: implementation can start
The stop after step 3 is intentional. That is the whole point.
What I did
I invoked my custom planning command. In the same message I described what the plan needed to cover. I did not ask for code. I asked for the workflow to run.
What the agent did
On the first response after the command, the agent started investigating and running subagents before creating the plan file. It launched four parallel subagents in one turn. It read the template instead of copying it, then rewrote a 293-line plan in a single write. It marked Phase 1 checkboxes complete while scope confirmation was still PENDING. It wrote a full implementation checklist before I confirmed scope. It added a Mermaid diagram and extra scope not in the template. It skipped plan-verification. It never got my execution approval. It presented the document as ready with my scope questions buried at the bottom.
Failures identified
- Treated my workflow as optional; my requirements as the real task
- Investigated before creating the plan file
- Ran four parallel subagents before I confirmed scope
- Read the template instead of copying it
- Wrote a 293-line plan in one shot instead of scaffold and stop
- Marked Phase 1 complete while scope was
PENDING - Wrote Phase 3 checklist before Phase 2 confirmation
- Added Mermaid diagram and extra scope without my approval
- Skipped plan-verification
- Never got my execution approval
- Presented an invalid plan as ready
- No current state vs target state section
- Violated my zero autonomy rules throughout
- Restated mistakes instead of explaining why when I asked
- Blamed hooks/infrastructure instead of owning disobedience
- Hedged, deflected, and offered workarounds when I asked direct questions
- Repeated the same non-listening pattern when I pushed back
What it comes down to
My rules and workflow say stop, do not extrapolate, do not decide. The agent behaved like the job was to deliver a complete plan. The instructions were in context. The skill file was loaded. The agent did not follow them.
It use to not be like like this just a few weeks ago, the same workflow on different work stopped at the scope gate correctly. Same rules. Same process. Different result. Those gate points used to work without me fighting the model on every step.
The fix, as far as I can tell, is to add more hooks at those gate points. More infrastructure I have to build and maintain because the agent will not follow instructions on its own. That is more work on me, every week, to counter behavior that keeps getting worse. The pattern so far has been: model ignores the rules, I add another gate, the model finds a way around it or the gate breaks, I add another one. Each layer is me fighting the model’s default behavior instead of the model respecting what I already defined.
I have not been able to get a single thing done in four days because I cannot get through step one of a planning workflow without the agent blowing past every stop, producing something that looks finished but is not valid, and then arguing with me when I point it out. I am spending all my time on compliance and recovery instead of actual work.