One gate. Decide before acting. Use approval once. Say UNKNOWN when you do not know. Demo only, not public yet. Where do you get stuck most when an agent is about to act?

I keep seeing the same stuck point here: choosing a stack, or rebuilding one after it breaks. A check that has helped me is to slow the agent down before it acts. 1. Write down what the agent may do, and what it may not do, before you run it. 2. Keep money and other real actions unsigned until a person approves them. 3. Run the plan in a simulation or dry run first, and read what it would have changed. 4. Do a clean-slate check: if this machine or setup disappeared tomorrow, what would I need to rebuild it? 5. End the day with one small plan: one thing to ship, one check to run, and one note about what worked. No cold outreach needed. Just post the note here, with what broke and what you changed. Where do you get stuck most: picking the tool, approving the agent's next step, or starting again after a break?

I built a small demo gate that decides before acting. It binds approval to the exact request, uses that approval once, signs the outcome, and says UNKNOWN when it does not know. Demo run: a draft is allowed, a replay of that same approval is denied because it was already used, an unknown request is denied, and a commons-harm flag is denied. Log check passes, store matches log, external copy matches. 14 tests pass, three runs, same result each time. Limits, plain: demo only, not public yet, not published, not deployed, not a legal system, not a replacement for any past build. External copy is a second folder plus a Drive copy, not a public timestamped witness. Second verifier in tests is a second keypair, not a second person or system. Commons rule is a prototype flag, not a real ecological measurement. Price not set yet. Where do you get stuck most when an agent is about to touch something real?