The Inference Report

The agent-focused repos are consolidating around a shared infrastructure problem: how to make agents persistent, portable, and interoperable across vendors. Munder-difflin and ai-memory both tackle the same friction point, agents need memory that survives between sessions and can hand off work to other agents without losing context. OpenViking goes further, treating memory, knowledge retrieval, and skill management as a unified database layer rather than separate concerns. These aren't competing solutions so much as different angles on the same bottleneck. Anthropic-Cybersecurity-Skills takes a different approach to the vendor lock-in problem by defining a standardized skill set across 29 security domains, mapped to six established frameworks. The bet here is that structured, portable skill definitions become more valuable than proprietary agent APIs, a framework that works with Claude Code, GitHub Copilot, and twenty other platforms simultaneously is solving for a world where agents are interchangeable but the work they do is not.

The broader pattern across discovery repos points to infrastructure maturation in two directions. On the data side, CVAT and Lightly Studio represent the unglamorous work of preparing training data at scale, annotation tools and multimodal data platforms that don't make headlines but determine whether models actually work. On the simulation and training side, Newton and the hands-on RL curriculum suggest researchers are moving past theory into tooling that bridges the gap between LLM alignment and practical agentic systems. Transformers remains the gravitational center, but the repos building around it are increasingly specialized: not "how do we train a model" but "how do we train it for robotics" or "how do we annotate video for it" or "how do we make agents that don't drift from their objectives." The knowledge base consolidation in Trilium and Project Tapestry hints at a parallel concern, whether the infrastructure should be centralized or distributed, whether models should be trained once globally or adapted locally. These aren't new questions, but the fact that they're being asked through code rather than conference papers suggests the field is past the point of theoretical debate.

Jack Ridley