1.6M agents registered for OpenClaw and did NOTHING.

The speaker opens by identifying a critical problem in the AI economy: 1.6 million agents were registered for OpenClaw at its peak, but most completed no tasks because users didn't know what to do with them. The core issue isn't a shortage of AI capability, but the inability to recognize which problems are agent-shaped and match them to appropriate solutions.

The speaker then presents academic grounding from Stanford's 2024 research and Anthropic's production systems. Stanford found that giving a cheap coding model 250 attempts instead of one increased bug-fixing success from 15.9% to 56%, and improvement followed a predictable law across four orders of magnitude. However, the critical finding was that when attempted over 10,000 times, correct answers existed in the pile over 95% of the time—but only if there was an automatic checker to validate them. Without mechanical verification (evals), performance plateaued at around 100 attempts.

Anthropić's research revealed that token spend explained 80% of the difference between successful and failed agent runs. Their multi-agent team beat frontier models working alone by 90.2% because teams can spend more tokens than a single agent can usefully hold. However, this only works with proper evaluation infrastructure.

The speaker identifies four questions to estimate in about one minute: (1) Size—Is the task larger than what one agent can hold in context? (2) Independence—Can parts be done without knowing what other parts did? (3) Separation of concerns—Do any parts need different minds or perspectives? (4) Checkability—Is verifying an answer much cheaper than producing one?

These lead to four possible verdicts: chat-based interaction for simple tasks, a single agent for tasks fitting in context that can check its own work, a team of agents for tasks requiring capacity or multiple perspectives, or no AI at all for judgment calls requiring human expertise.

The speaker demonstrates with three real examples. First, scheduling a gym slot fits a context window and is a straightforward single agent task. Second, reviewing 40 tools' contracts, emails, and usage logs across thousands of pages requires a multi-agent system with proper evaluation—the agent team must verify sources match tasks or reject entries, retry failures, and maintain a scorecard. Third, hiring decisions and product direction require human judgment; AI should support research but not replace expert human instinct.

The speaker emphasizes that judgment call problems are where people most often over-delegate to AI. No frontier model beats an expert in their domain. The cheapest solution is sometimes sitting down and applying your own judgment.

Finally, the speaker describes a tool that implements this one-minute test, allowing users to describe their task, set sliders for four estimates plus cost factors, and receive a verdict with next steps attached—whether that's clicking into a multi-agent Ringer setup, ChatGPT, or a reminder to use human judgment instead.