Coding as the epicenter of AI progress and the path to general agents (www.interconnects.ai)

OpenAI’s new GPT‑5-Codex and the rise of CLI coding agents were highlighted as concrete signs that coding is the next tractable, broadly useful frontier for frontier models. OpenAI reported GPT‑5 (with an ensemble and an experimental reasoning model) outscored humans at the ICPC, and GPT‑5‑Codex specifically is tuned for code: it dynamically allocates “thinking” time based on task complexity, pairs interactively with developers, and can persistently execute and iterate on long-running tasks (tests showed >7 hours of independent work). Practically, agents now do more than autocomplete — they run tests, fix git, call local tools and web search, and handle mini-projects that used to take days in hours. Historical trajectory: function completion (~2021), scripting (~2022), CLI agents for small projects (~2025), with complex production codebases estimated later. The significance is twofold: product and prompting matter as much as raw model scale, and coding agents provide the clearest path to general, persistent assistants. Performance gaps often come from scaffolding and product design (same base model can behave very differently across implementations), and calibration — avoiding “overthinking” — is crucial. Adoption is slower because CLI agents change workflows, distribution channels, and UX, but they lower entry barriers, enable new agent-driven workflows (editorial assistants, repo-wide searches over hundreds of thousands of tokens), and point toward general agents that combine interactive pairing with long-horizon autonomous execution.