The Inference Report

The infrastructure for AI agents is solidifying around two distinct needs: control and observation. On the control side, projects like CUA and openpilot are building the sandboxes and SDKs that let agents actually do things, CUA explicitly targets desktop automation across operating systems, while openpilot demonstrates the pattern at scale by controlling 300+ cars. Smaller runtimes like KrillClaw (49KB in Zig) show the opposite impulse: proving that agent execution doesn't require framework bloat. On the observation side, TensorZero and the MCP Inspector are addressing a real gap in the stack. You can build an agent now, but running it in production without visibility into what it's doing or how well it's working is still painful. TensorZero unifies the gateway, observability, and experimentation in one place, solving the problem that PostHog solves for product analytics but specifically for LLM pipelines. The MCP Inspector takes a narrower angle, focusing on debugging the protocol layer itself.

Skills and memory are becoming first-class concerns. GitNexus solves a specific problem: drop a codebase in your browser and get a queryable knowledge graph without spinning up infrastructure. Beads does something similar for agent context, it's explicitly framed as memory augmentation for coding agents. Skills repos like the Composio curated list and mattpocock's Skills collection are treating agent capabilities as composable, versioned artifacts rather than one-off integrations. This mirrors how package ecosystems matured: the infrastructure works, now the question is how you organize what the infrastructure can do. The surge in agent-specific tooling suggests the bottleneck has shifted from "can we build agents" to "can we make them reliable, observable, and useful at scale."

Jack Ridley