The Inference Report

The agentic framework wave is real, and it's fragmenting into specialized layers. Obra's superpowers and LangChain's deepagents are both hitting the charts because they solve a concrete problem: agents that need to reason, plan, and spawn subtasks fail badly without scaffolding. Superpowers packages the methodology alongside the code. Deepagents gives you the harness, planning tools, filesystem backends, subagent spawning, and lets you wire your own logic. These aren't competing so much as targeting different points on the abstraction spectrum. The friction point both address is that naive agent loops either hallucinate or get stuck; structured orchestration prevents that. CopilotKit's AG-UI Protocol suggests the next layer: if agents are going to be common, the UI layer needs to understand them natively. Claude HUD's dashboard approach, showing context usage, active tools, running agents, indicates developers want visibility into what their agents are actually doing, not just the final output.

Meanwhile, the supporting infrastructure is solidifying around knowledge representation and retrieval. Haystack's explicit control over retrieval, routing, and memory reflects a maturing understanding that RAG and agent systems need deterministic pipelines, not black boxes. GitNexus takes a different angle: client-side knowledge graphs that run in the browser eliminate the server dependency entirely, trading latency for privacy. SDV's focus on synthetic tabular data and Weights & Biases' production model management suggest teams are moving past experimentation and into the operational phase, they need to test on realistic data and track what's running where. The discovery of Whisper on Android and NaViT's vision transformer work indicates agents aren't just text-based anymore; multimodal reasoning and on-device inference are becoming table stakes.

Jack Ridley