Context is the bottleneck for coding agents now (runnercode.com)

OpenAI’s recent perfect score on the 2025 ICPC using a GPT-5 variant shows frontier models now have superhuman raw programming ability for closed‑world problems where every requirement is stated. But the article argues that the real bottleneck for practical coding agents isn’t intelligence anymore — it’s context. On a spectrum of autonomy (from single-line autocomplete up to owning an entire codebase), current agents reliably operate around “one commit” level on real production codebases and fail to move past larger tasks because they lack the rich, distributed context human engineers use. That context goes far beyond files and docs (which many agents can already access) and includes high‑level codebase organization, emergent architectural patterns, historical rationale (why things were implemented a certain way), deployment/testing conventions, and product or regulatory requirements often buried in PRs, incident post‑mortems, Slack threads and tribal knowledge. Solving this requires more than connectors: agents need sophisticated preprocessing, retrieval and synthesis across heterogeneous sources, provenance and uncertainty estimation, long‑term memory, and the ability to detect missing context and ask humans targeted questions. Practically, that means human-in-the-loop workflows will remain essential for the foreseeable future, and research should shift from raw model capability to infrastructure and algorithms for assembling, summarizing and reasoning about distributed contextual knowledge.