Exact-match scoring fails when evaluating natural language predictions against structured AI agent actions. This guide explores building a semantic LLM judge to bridge the gap between human guesses and tool-based outcomes, using a multi-component rubric for tool, action, and reasoning matches.
8 min read min read
The Signal
Platform intelligence, agent economics, and the infrastructure shaping autonomous AI.
Payments, protocols, and the plumbing that makes machine-to-machine commerce work.
Market timing, competitive dynamics, and what the data says about agent adoption.
Building better agents. Soul design, composition patterns, and quality measurement.
AI agents currently face an accountability gap when performing tasks like phone calls or sales recovery. Cryptographic receipts solve this by providing signed, content-addressed records of execution that integrate with USDC settlement on Base to ensure agents actually did their jobs.
J Nicolas
•
•
8 min read min read
Building a production voice AI agent requires balancing latency, accuracy, and reliability. This guide explores the five-stage audio pipeline, the 800ms latency budget, and architectural patterns for handling real-time telephony using the OneShot SDK.
J Nicolas
•
•
7 min read min read
Most AI agents are stateless functions with no reputation. soul.md is a structured markdown file that defines an agent's values, expertise, and services, allowing them to function as persistent economic actors with consistent personalities and ethical constraints across tasks.
J Nicolas
•
•
8 min read min read
Most AI agent projects fail because they skip the measurement layer. By following the six levels of the SolveEverything framework—from Ill-Posed to Commoditized—teams can move past 'vibes-based' evaluation and build industrialized, reliable agent commerce systems.
J Nicolas
•
•
8 min read min read
Agent commerce introduces unique security risks when AI agents transact at machine speed. This guide explores how to defend against replay attacks, runaway agent spending, and price exploitation using nonces, tiered rate limiting, and circuit breakers.
J Nicolas
•
•
7 min read min read
Most e-commerce stores fail at cart recovery by waiting too long. Freway uses AI to detect buyer hesitation in real-time, treating the checkout process as a conversation rather than a form to guide uncertain shoppers toward a completed purchase.
J Nicolas
•
•
7 min read min read
Traditional SaaS and API billing models fail in the age of AI agents. To scale autonomous commerce, we must shift from paying for effort and API calls to paying for verified outcomes and successful tasks using protocols like x402.
J Nicolas
•
•
7 min read min read
Moving beyond simple demos requires agent composition. By treating AI agents as single-capability primitives with defined schemas, developers can build complex, reliable production workflows using patterns like pipelines to coordinate research, drafting, and verification tasks.
J Nicolas
•
•
8 min read min read
In the race to build agent marketplaces, features are easily copied. The real competitive advantage lies in creating an authoritative measurement system—a scoreboard—that defines quality, attracts builders, and earns the trust of buyers through standardized performance benchmarks.
J Nicolas
•
•
6 min read min read
AI is accelerating the industrialization of professional domains from decades to years. By tracking shifts from effort-based pricing to outcome-based models and the rise of machine-verifiable proofs, we can identify which industries are about to be absorbed into AI infrastructure.
J Nicolas
•
•
8 min read min read
While AI models can reason, they lack the infrastructure to pay, discover, and verify identity. The next 18 months represent a critical window to establish the standards for agent-to-agent commerce before legacy patterns calcify, much like DNS and payment rails of the past.
J Nicolas
•
•
7 min read min read