LazyCop: Steer IBM Bob While It Works

THE PROBLEM: YOU FIND THE MISTAKE TOO LATE: You give an AI agent a task and wait. Early on it guesses wrong, and every step after builds on that bad guess. You only see it when you read hundreds of finished lines. That waiting is expensive. One IBM Bob user watched a simple formatting fix burn 150 Bobcoins (about $75), 542 searches and 61 minutes before stopping it by hand (ibm-bob issue #3683). The real problem isn't that agents make mistakes. It's that we catch them after they're paid for. THE IDEA: REVIEW WHILE BOB WORKS, NOT AFTER: LazyCop is the cop on duty while you wait. It turns dead time into a short, two-way quiz. Before Bob edits code, it says what it's about to do and what it's assuming. LazyCop shows that as a card: "Bob thinks an expired coupon means its timestamp is before now. Do you?" Bob's answer is one option. Next to it are two or three other ways to read the task, plus "Something else" and "Ask Bob why." One tap is enough. - Agree, and you both know you're on the same page. - Pick another answer, and Bob gets your correction at its very next step. No restart, no lost work. - Ask why, and Bob explains before it writes a line. There's no answer key and no way to be "wrong." Every disagreement is a win: either you learn something about your own code, or you catch a mistake before it costs more. WHO IT'S FOR AND HOW IT FEELS: LazyCop is for any developer who hands real tasks to IBM Bob. A panel next to Bob's chat shows the files Bob is visiting and a live Bobcoin counter against your budget. Cards show up one at a time, never block your screen, and skipping costs nothing. Bob only waits when it matters: never for small things, a few seconds for an assumption, and a bit longer for big decisions. WHY IT'S DIFFERENT Most tools review code after the agent is done. LazyCop moves review into the run, and makes it feel like a game instead of a chore. It doesn't grade you or Bob. It checks one thing: do you and your agent understand the task the same way? Every card comes from Bob's own words: what it declared, what its diff does, what it said at the end. Nothing is made up. LazyCop also finds "hidden assumptions," rules Bob quietly applied but never said out loud. In our tests, those mattered most. IT ALREADY WORKS LazyCop sits next to Bob, not inside it. Bob's own hooks report each step, and a small LazyCop MCP server carries your answers back into the running task. In our test on Bob 2.2.0, a one-tap correction sent from a browser page reached Bob mid-task. Bob replied, updated its assumption and rewrote its logic in 66 seconds, with no rework. Less waiting. Fewer wasted coins. Code you understand before you merge.