Trust scores for AI agents

A customer service agent promises 95% accuracy. A coding agent says it can handle edge cases. What developers actually ship is demos, not documentation. Enterprises are purchasing faster than anyone can validate what actually performs, and the AI agent market offers them zero incentive to pump the brakes. One company rolled out an unvetted agent last quarter. It hallucinated pricing information to 400 customers before someone finally noticed. Crucible validates what demos never could. Developers submit their agents for standardized evaluation against a published rubric. Every agent receives a public scorecard: accuracy benchmarks, documented failure modes, uptime monitored over 30 days, and verified feedback from teams using it in production environments. Procurement teams weigh evidence rather than sales pitches. As the library expands, patterns emerge. Which architectures deteriorate under heavy load. Which developers actually push fixes when things go wrong. Which agents still perform when real money is at stake. The starting point is customer service agents, where enterprise adoption is most mature and failure costs translate directly into dollars. Evaluation criteria center on what procurement teams genuinely want answered before signing: accuracy against real-world scenarios, response time under stress, and how the developer reacts when something breaks. The plan is to test the 20 most prominent agents and publish scorecards through a free directory built on Airtable and Webflow. Those initial 20 serve simultaneously as proof of concept and marketing engine. Developers promote strong scores. Procurement teams save the site. AI media outlets report on the rankings. Five enterprise design partners help shape the criteria in return for early access. Whatever they require on a scorecard before greenlighting a vendor dictates what gets measured. Developers pay $149-499 for verification and re-verification as agents get updated. Enterprises pay $500-2,000/month for API access to scorecard data plus compliance-filtered search capabilities. Two revenue streams drawn from opposite ends of the same trust deficit. Today, procurement teams are making buying decisions based on demos and optimism. Crucible inserts itself as the critical layer between the sales pitch and the purchase order.

What you get with the full dossier