The Inference Report

The GitHub ecosystem is consolidating around two parallel movements: infrastructure for AI agents that can persist and coordinate work across sessions, and tooling that lets those agents operate more efficiently within existing developer environments.

The first trend shows up across TencentDB-Agent-Memory, loopx, and code-review-graph. These repos solve a concrete problem that emerged once developers started shipping multi-step AI workflows. A single conversation with an LLM produces useful output, but agents that need to maintain goals across days, handle quota constraints, or hand off work to other agents require durable state. TencentDB-Agent-Memory turns conversations and code into reusable memory assets governed across frameworks. loopx adds executable todos and evidence logs for long-running teams. code-review-graph builds a persistent map of the codebase so subsequent agent queries don't waste tokens re-reading the same files. These aren't glamorous infrastructure problems, but they're the ones preventing agent systems from becoming expensive toys.

The second trend is about giving agents better access to developer tools without forcing them to learn entirely new abstractions. Cloudflare's computer lets agents control a machine programmatically. firecrawl's pdf-inspector handles a specific, common task well: it detects whether a PDF is scanned or text-based before deciding how to extract data, reducing wasted inference on OCR when it isn't needed. DeepSeek-Reasonix and AutoGPT remain high-signal repos because they're proving that agents work best when engineered for their specific constraints, not when layered with generic frameworks. The skill-set repos (agent-skills, mattpocock's skills, obra's superpowers) are trending because they let teams package their own domain knowledge as reusable units that agents can call, rather than forcing every team to teach the same lessons to every model. This is the inverse of the 2023 pattern of "everything should be a prompt." What's working now is treating agent capabilities as first-class code artifacts that can be versioned, tested, and shared like any other library.

Jack Ridley