The Inference Report

The GitHub conversation around agents has shifted from "what if we built this" to "how do we actually run this." Ruflo and agency-agents occupy the high-star territory of multi-agent orchestration platforms, but the real traction is in the narrower, more specific tools: terminal-based coding agents like DeepSeek-TUI and Reasonix that don't pretend to solve everything, context window optimizers like context-mode that measure their gains in concrete percentages, and research-focused agents like dexter and local-deep-research that solve a single problem well. The pattern is clear. Developers are less interested in frameworks that require buying into a complete worldview and more interested in tools that integrate into existing workflows. Dify's 140k stars reflects this shift toward production-ready platforms that let you wire agents together without reinventing infrastructure, while the smaller discovery repos like Kiln and TransformerLab suggest people are also investing heavily in the evaluation and optimization layer that makes agents actually usable.

The supporting infrastructure tells the story better than the agents themselves. Context-mode's 98% token reduction and local-deep-research's support for 10+ search engines indicate developers are solving the real bottleneck: not building agents, but making them cheap enough and accurate enough to deploy. Label-studio and the synthetic data tools in the discovery section show the pipeline: you need clean data to train or fine-tune anything, you need evals to know if it works, and you need memory systems like PowerMem to make agents retain anything useful across sessions. The presence of TabPFN, a foundation model for tabular data, and the various multi-model databases suggests another quiet shift: agents aren't just language models anymore. They're becoming orchestrators that coordinate different model types and data sources. Meanwhile, coding-interview-university's 346k stars and Karpathy's CLAUDE.md file (115k stars) sit at the top because developers still want to understand what they're building, not just plug in magic. That's where the real signal is.

Jack Ridley