Tech • AI • Robotics • Game
A Staff Engineer Collaborator” | Notion’s First Look at GPT-6 Astra
7/10TL;DR
Notion is using an AI coding model called Astra as a high-level engineering collaborator, relying on it to investigate costly token usage, uncover missed optimization patterns, and speed up back-and-forth problem solving.
KEY POINTS
AI positioned as a staff-level collaborator
At Notion, the model is being framed less as a tool for small, tightly scoped tickets and more as a collaborator for difficult engineering problems. The approach is to assign a hard objective that has already resisted earlier attempts, mirroring how teams might engage a staff engineer on a complex systems issue.
Focus on growth and abuse prevention
Recent work has centered on balancing strong AI product experiences with practical controls around cost and misuse. One key challenge has been making services useful to customers while also reducing avenues for users to obtain free tokens or otherwise exploit the system.
Clearer communication than earlier models
Engineers described Astra as notably better at communicating findings than earlier systems such as Sora and o1. Previous models were seen as capable of identifying interesting issues but weaker at explaining what they had found, while Astra produced responses that were easier to understand on a first read and closer to how a human engineer would summarize a problem.
Shift from task execution to goal-driven work
The coding workflow has changed from assigning discrete tasks to providing broader goals and helping the model navigate toward them. That has shortened the amount of time engineers spend actively interpreting the model’s questions, tightening the feedback loop and making collaboration faster.
Token costs and cache reuse as a core engineering problem
A major early assignment for Astra involved token spending, a significant cost center for AI systems. The investigation focused on cache reuse, including whether existing systems were performing as close as possible to their theoretical maximum and whether prior optimization work had overlooked inefficiencies.
Missed patterns surfaced in prior audits
Engineers provided Astra with previous work and the current understanding of the system, then asked what had been missed in earlier audits. The model identified patterns that had gone unnoticed, connecting clues more quickly and pointing to scenarios that implied broader underlying issues in the infrastructure.
Concise findings improved decision-making
The usefulness of the model was not only in pattern recognition but in the precision of its explanations. Results were described as crisp and free of unnecessary detail, which made it easier for engineers to absorb the diagnosis quickly and act on it without repeatedly parsing long outputs.
Recommended prompting style: give outcomes, not instructions
Internally, the guidance for teammates is to provide a topline objective instead of step-by-step tasks. One example of the desired scope was to point to a dashboard metric and say that it needs “one more nine,” a shorthand for pushing reliability or performance to another order of magnitude rather than asking for a narrow fix.
CONCLUSION
The emerging use case at Notion suggests AI coding systems are becoming more valuable when treated as strategic engineering partners rather than automated ticket workers. Their impact appears strongest where costs, infrastructure efficiency, and hard-to-spot system patterns matter most.
Ask a question