Impressions from visiting OpenAI, Anthropic, & Cursor
Source
Gergely Orosz
Author
Gergely Orosz
Date
Key takeaways · AI-distilled
At OpenAI, more than 95% of non-engineers use Codex rather than ChatGPT. The coding agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →, not the chatbot, is becoming the general-purpose interface for people who do not write code.
Cloud agents have no way to complain. A local AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → surfaces warnings to the human watching it; a long cloud run cannot. Cursor's workaround is to interview the model afterward so it "confesses", then give those confessions to the infra team.
OpenAI bought Ona, formerly Gitpod, because cloud development environments turn out to be the right primitive for agents: secure, persistent sandboxes that outlive one session or one laptop.
Why now: models only became good enough to code unattended around Opus 4.5 and GPT-5.4, context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → windows reached ~1M tokens, MCPThe Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.Full definition → and skills matured, and cloud providers finally have spare GPU capacity.
What makes a Slack mention feel like a paradigm shift is not Slack. It removes the per-machine MCP and skills setup and the tool switch, so one typed line starts a run that does not live on your laptop.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
Why it matters
Firsthand reporting from inside OpenAI, Anthropic, and Cursor reveals where coding-agent workflows are heading — cloud agents, agent-environment engineering, and aggressive per-token cost optimization — so you can anticipate how your own stack and role will shift.
Key quotes
“At OpenAI, more than 95% of non-engineers use Codex, not ChatGPT.”
“As Codex becomes more capable, its most valuable work is unfolding over hours or days, rather than minutes.”
