The Inference Report

The clustering here reveals two distinct developer preoccupations running in parallel. One camp is building agentic development environments and agent frameworks, Warp, Superpowers, and the various agent harnesses, treating the developer as someone who should delegate routine work to autonomous systems. These repos treat the terminal and IDE as platforms for orchestrating AI labor rather than places where humans write code directly. The other camp is solving the practical infrastructure problems those agents create: how to give them memory, how to route their API calls, how to track what they're doing, how to keep them from leaking sensitive data. Presidio detects PII before it reaches an LLM. The memory services and knowledge graph builders exist because agents without persistent context are expensive and unreliable. This isn't hype, it's the shape of real operational friction.

What's notable is the absence of consensus on what "agent framework" means. Superpowers, Codex skills, the Coding Agent Harness, and GitNexus all claim to solve agent orchestration but describe completely different architectures. Some run client-side, some run as middleware, some as CLI tools. The diversity suggests the category is still contested, which means early adopters are paying the integration cost. Meanwhile the infrastructure plays, Unsloth for local model training, Pixeltable for multimodal data pipelines, TransformerLab for model evaluation, are gaining traction because they solve problems that don't require buying into a specific agent philosophy. They're tools for the people building the tools. The real momentum isn't in the agent frameworks themselves but in the supporting layer that makes agents feasible to run, debug, and reason about at all.

Jack Ridley