What the Heck Is Graph Engineering?

OpenAI Paused Astra After Cyber Risk Findings
- OpenAI paused its Atlas/Astra model after internal tests showed advanced agentic cyber capabilities that risked real-world exploits.
- Nathaniel describes OpenAI isolating testing, encrypting weights, and expanding sandbox monitoring before any public release. (Time 0:00:37)
ByteDance Is Training A Potential 10T Parameter Model
- ByteDance reportedly began a pretraining run for a base model that could reach ~10 trillion parameters, potentially matching frontier scale.
- Nathaniel notes the run could take 3–6 months plus RL training and ByteDance hasn’t finalized release size. (Time 0:03:55)
Open Weights Meet Revenue Sharing Experiments
- Some model publishers are exploring revenue sharing for published weights to capture commercial upside while keeping models broadly usable.
- Nathaniel cites Moonshot’s reported 30% deals and Reuters’ report about Alibaba considering revenue sharing for Quen. (Time 0:07:37)
Use AutoMode To Reduce Permission Friction Safely
- Adopt AutoMode for coding agents to reduce prompt friction while relying on classifiers to block destructive actions.
- Nathaniel reports Anthropic found AutoMode caught 89% of harmful actions versus 13.6% by human reviewers. (Time 0:09:10)
Engineering Evolved From Prompts To Harnesses
- The evolution from prompt to context to harness engineering shows increasing focus on surrounding models with information and tools.
- Nathaniel highlights context engineering as controlling what the model sees and harness engineering as shaping the environment and tools. (Time 0:11:48)
Loops Make Single Agent Behavior Programmable
- Loops program individual agent behavior: observe, plan, act, check, and repeat until a measurable stop condition.
- Nathaniel emphasizes loops are best for single-agent tasks with clear finish lines and measurable outcomes. (Time 0:19:52)
Graph Engineering Organizes Multiagent Systems
- Graph engineering designs how multiple agents, tools, knowledge sources, and humans connect and hand off work.
- Nathaniel explains each graph node runs its own loop and edges define data flows, dependencies, and failure handling. (Time 0:20:57)
Choose Between Org Graphs And Work Graphs
- Distinguish org graphs from work graphs: org graphs are long-lived with persistent agents; work graphs are dynamic and ephemeral for specific tasks.
- Nathaniel gives an org-graph example automating research-to-publish workflows that recur daily. (Time 0:23:44)
Only Build Graphs When Single Loops Fail
- Don’t build complex multi-agent graphs unless your process needs parallelism, specialty handoffs, or resilience to single-node failure.
- Nathaniel advises using single loops when one agent with a clear finish line and context window suffices. (Time 0:24:52)