The Inference Report

The trend here splits cleanly into two camps: agentic coding tools gaining real traction, and infrastructure quietly getting its house in order. The coding agent space is crowded and consolidating around a specific pattern. OpenAI's Codex, Anthropic's Claude Code, and the skill frameworks built atop them (obra/superpowers, affaan-m/ECC, multica-ai/andrej-karpathy-skills) all solve the same problem: how to make LLMs useful at actual development work without requiring you to context-switch out of your terminal. The differentiation isn't in the agent itself anymore but in the skill harness, the memory layer, and the abstractions that let you compose routine tasks into workflows. That these repositories are accumulating hundreds of thousands of stars suggests developers have moved past asking whether agentic coding works and are now asking which harness lets them control it best. The appearance of mattpocock/skills and multica-ai/andrej-karpathy-skills alongside the tools themselves signals that the real value is migrating away from the agent framework and into the configuration layer, the prompt engineering, the institutional knowledge about what actually breaks LLM code generation.

Meanwhile, the infrastructure tier is doing necessary work that doesn't trend as visibly. n8n-io/n8n combines workflow automation with native AI capabilities and 400+ integrations, solving the integration tax that kills most internal tool projects. PostHog has broadened from analytics into observability and error tracking, essentially building the diagnostic layer that agents need to understand whether they've made things worse. Makeplane offers an open-source alternative to Linear and Jira, which matters less for its features than for what it signals: developers want to own their tooling and their data, not rent them. The discovery repos point toward harder problems being tackled quietly. ONNX remains the only serious attempt at model interoperability. IsaacLab addresses robot learning, a domain where simulation and real hardware divergence is still unsolved. The model merging and modality-missing learning collections suggest the field is moving past the era of "bigger model, better results" and into the era of composition, where the interesting work is figuring out how to combine partial solutions into something coherent. None of this is as visible as a new coding agent, but it's the infrastructure that will determine whether agentic development becomes reliable enough to ship to production.

Jack Ridley