The Inference Report

The agent skills ecosystem is solidifying around a few competing abstractions. Repos like obra/superpowers, anthropics/skills, and mattpocock/skills all position themselves as frameworks for defining reusable agent capabilities, but they're solving slightly different problems. Superpowers frames itself as a methodology for software development; anthropic's skills emphasize public, shareable capabilities; mattpocock's collection reads as personal tooling made public. What's notable is that none of these are fighting for dominance through features alone. Instead, they're winning by integrating with existing developer environments: Claude Desktop, Cursor, Windsurf, n8n. The MCP protocol has become the connective tissue, and repos like czlonkowski/n8n-mcp show developers building bridges between their workflow tools and AI-powered automation. This suggests the real value isn't in the framework itself but in how quickly it plugs into the tools developers already use.

Infrastructure for observing and controlling AI systems is gaining ground alongside the skills layer. Future-agi positions itself as an end-to-end platform for tracing, evals, and simulations; wandb/wandb and screenpipe/screenpipe take different angles on the same problem of visibility into AI behavior. Screenpipe records everything locally; wandb focuses on training and model lifecycle; future-agi emphasizes safety through guardrails and observability. Meanwhile, performance tooling like oven-sh/bun and supertone-inc/supertonic address practical constraints: bun as a unified JavaScript runtime and toolchain, supertonic as on-device multilingual TTS that doesn't require cloud calls. These aren't flashy additions to the ecosystem. They're solving the unglamorous work of making systems faster and cheaper to run. That combination, skills plus observability plus performance, reflects where engineering effort is actually flowing: toward making AI agents practical rather than impressive.

Jack Ridley