The Inference Report

The terminal and agentic development environments are consolidating around a new abstraction layer. Warp and Ghostty compete on the infrastructure side, one through an agent-first terminal, the other through a GPU-accelerated emulator, but both signal that developers want their tools to do more than passively wait for input. The real momentum, though, sits in the frameworks building on top of that foundation. Obra's superpowers framework has accumulated significant traction by positioning itself as both a skills library and a development methodology, while browserbase's Claude Agent SDK takes a narrower approach by bundling web browsing into a specific agent harness. These aren't competing so much as staking out different surfaces of the same problem: how do you compose agent capabilities in a way that doesn't require rewriting the entire stack for each new task.

The secondary trend is data infrastructure catching up to agent demands. Dagster's orchestration platform and FLAML's AutoML tuning library address a practical constraint: agents generate work, but someone still needs to schedule it, monitor it, and tune it reliably. Local Deep Research achieving 95 percent on SimpleQA while supporting multiple LLM backends suggests developers want to run research agents without vendor lock-in, which is distinct from the cloud-first agent frameworks that dominated earlier cycles. The collection of public APIs and the various agent harnesses (jcode, maigret) point to a pattern where the bottleneck has shifted from "can we call this service" to "can we coordinate calls across services without hallucinating the wrong endpoint." That's a different problem than building the agent itself, and it's where the friction actually lives in production systems.

Jack Ridley