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

OpenAI’s frontier models (a high‑compute GPT‑5) recently scored perfect on the 2025 ICPC, underscoring that raw model intelligence and programming knowledge are reaching or exceeding human levels on well‑scoped problems. Yet practical coding agents still can’t replace developers because the bottleneck isn’t “IQ” but context: current agents reliably handle small tasks (autocomplete to a single commit—Level 2 autonomy) but struggle with larger work like PRs, major features, or evolving codebases (Levels 3–5). Competition problems provide all needed context in a single prompt, which is why models shine there; real engineering requires far more scattered, implicit, and historical context. That context includes basics (access to all source files, docs, and runnable tests) plus subtler, distributed knowledge: repository organization, emergent architectural patterns, rationale for past decisions held in commits, incident postmortems, Slack threads, CI/CD quirks, and business/regulatory constraints. Much of this isn’t in one file and must be synthesized, not merely read—agents today have maybe 20% of the relevant context and often plow ahead instead of asking for missing info. Moving agents forward will require broad access plus sophisticated preprocessing, long‑range retrieval/memory, cross‑source synthesis, and uncertainty detection so they can request human guidance; until then experienced developers remain necessary to fill the unwritten gaps.