Your first agent will not fail because the model is weak. It will fail because nobody decided what it was allowed to do. On 30 September 2026, Gartner reported that 54 per cent of organisations have no defined approach to limit AI agent access, or simply give agents the access a human would have. That came from a survey of 297 cybersecurity leaders in the second quarter of 2026. Gartner's advice was to govern agents by what they can do, not by how clever the model is. If large companies with security teams are getting this wrong, solo builders should assume we are too. I trained as an auditor at PwC and KPMG. When I review any system, human or AI, I ask the same five questions. Run them on your agent before you switch it on: 1️⃣ What can it touch? List every file, account and API key it can reach. Remove anything the task does not need. 2️⃣ What can it do without asking you? Drafting an email is low risk. Sending it, spending money or deleting data is not. Put an approval step in front of anything you cannot undo. 3️⃣ Where is the log? If you cannot see what the agent did yesterday, you cannot fix what it gets wrong tomorrow. 4️⃣ How do you stop it? Set a spending cap and know where the off switch is before you need it. 5️⃣ Who checks the output? Review a sample of its work every week. Ten items is enough to spot drift early. This is the Controls and Trust part of the ACCOUNT Framework from my book, Human-Led, AI-Powered Finance. I wrote it for finance teams, but it applies to any agent that touches money, customers or data. Build fast. Just decide the limits before the agent does. Which of the five is missing from your current build? Drop the number below and I will share how I would close it.