Summary
AI Jason cuts through the Twitter confusion around ‘graph engineering’ by separating it into three distinct concepts that practitioners have been conflating: control graphs (structured SOPs that make agent behavior reliable), knowledge graphs (entity relationship stores for better retrieval), and graph of loops (an emerging framework for managing multiple continuously improving autonomous loops). The video focuses most of its depth on control graphs, which it argues is what influential voices like Peter are actually advocating for.
The practical core of the video covers two implementation paths. The first is code-based control flow using tools like LangGraph, Claude Code’s dynamic workflows, or pre-written Bash scripts with deterministic checkpoints. The second — and more accessible — approach uses the LLM itself as the graph engine: modern frontier models can follow complex multi-step SOPs written in plain text, JSON, or Mermaid diagrams well enough that explicit graph code is often unnecessary. Real examples include an autonomous ML training loop that tests changes, checks a metrics file, and reverts failures without human prompting, and a deep research agent that fans out sub-agents and self-evaluates completeness before generating output.
A key strategic insight is the shift happening at AI-native companies from ‘human prompting the agent’ to ‘something else prompting the agent on behalf of the human’ — specifically event-driven loops and automated triggers. Claude Code, Codex, and similar harnesses already provide the primitives (bash, sub-agents, hooks) to implement these patterns today, and the video argues the gap between teams that exploit these features versus those that don’t is already significant.
📺 Source: AI Jason · Published August 25, 2026
🏷️ Format: Deep Dive
