The short version

Most mistakes builders make are not technical. They are structural. Builders give agents more access than they can control, skip checking the output, and assume that a higher level always means better results.

At every level

These three mistakes happen regardless of which tool you use.

Treating output as truth
The agent generates text that sounds right. Not facts that are verified. Always check the output before you use it.
Missing the access question
Most people focus on what the agent can do. The real question is what it can reach. Access and capability are different things.
Adding tools without adding understanding
Every step up changes what the agent can affect on your system. New tools require new mental models, not just a new subscription.

Mistakes by level

APL 1 Browser chat
  • Pasting private data into chat. Your prompts may be stored or used for model training.
  • Staying too long. If you copy-paste between chat and your work 20 times a day, you need APL 2.
  • Not learning prompt patterns. Better prompts produce dramatically better output at every level.
APL 2 In-editor assistant
  • Accepting suggestions without reading. Cursor and Copilot make it easy to click Accept too fast. The human in the loop only works if the human actually reads.
  • Assuming the agent sees your full project. Most APL 2 tools see the current file, not the full architecture.
  • Skipping version control. If you do not commit before agent edits, you cannot undo cleanly.
APL 3 Full file access
  • Not reviewing all changed files. The agent modified 12 files. Did you check all of them? At APL 3, staying human in the loop requires active effort.
  • Giving broad permissions "just for now." Temporary always becomes permanent.
  • Skipping tests. The agent wrote it is not a testing strategy.
  • Losing context across sessions. Long sessions cause the agent to forget earlier changes.
APL 4 Multi-agent systems
  • No logging. Multi-agent systems need records at every boundary.
  • No way to stop the pipeline. If you cannot stop it, do not start it.
  • Trusting agent coordination. Agents share data. They do not share understanding. Those are different things.
  • Building for scale before correctness. Get one agent right before running four.

The root cause

PSA Formula
1 Problem What are you trying to solve?
2 Setup What tool level fits that problem?
3 Access What access does that setup need?

Most builders do it backwards. They buy the tool first, then look for things to do with it. The right order is the PSA Formula: start with the Problem, choose the Setup it needs, and only then grant the Access that Setup requires. Access without a clear problem has no clear boundary.

Higher is not better. The right level is the one where your control matches your access. A well-governed APL 2 setup is safer than an ungoverned APL 4 one.

Not sure where you stand? The AI Setup Snapshot finds your level in 3 minutes and shows you exactly what to fix.

Take the AI Setup Snapshot