Coding Agents Don’t Lack IQ—They Lack Context

The author argues that coding agents fail mainly due to missing context, not insufficient intelligence. While models excel on contest problems, real-world development depends on implicit knowledge about architecture, history, practices, and business needs that isn’t centrally documented. Progress requires richer contextual ingestion and synthesis, human guidance for gaps, and agents that know when to ask for help.

Key Points

Model intelligence is increasingly sufficient; the limiting factor for coding agents is missing real-world context, not raw capability.
Current agents reliably handle tasks up to about one commit (Level 2) on existing codebases; larger autonomous scopes fail primarily due to context gaps.
Essential context includes not just code and docs but emergent architecture, historical decisions, unwritten dev/deploy practices, and product/business requirements.
This context is fragmented, often undocumented, and requires sophisticated preprocessing and synthesis—simple file or tool access is not enough.
To progress, agents must be fed richer context, keep humans in the loop to fill inevitable gaps, and learn to detect and ask for missing context.

Sentiment

The community is broadly sympathetic to the general observation that context matters for agent effectiveness, but significantly divided on whether it is truly the primary bottleneck. A substantial contingent argues that intelligence failures, context poisoning, and architectural limitations of transformers are more fundamental problems than context scarcity. The discussion is notably pragmatic, with many commenters sharing concrete workarounds and tool recommendations rather than engaging in purely theoretical debate. There is a healthy undercurrent of skepticism toward AI hype, with several commenters questioning whether current models genuinely represent intelligence improvements or just benchmark optimization. Overall, the tone is one of cautious, experienced practitioners sharing hard-won insights rather than either uncritical enthusiasm or blanket dismissal.

In Agreement

Context is indeed a critical limitation, and agents struggle spectacularly with large codebases and files that exceed their context windows, particularly in languages like C++
Well-structured, well-documented codebases with smaller files produce dramatically better agent output, confirming that context quality matters enormously
Agents need hierarchical, summarized context rather than raw code dumps -- similar to how humans use compressed mental models rather than memorizing entire codebases
Sub-agents with isolated context windows are a promising architectural pattern for managing context, as they prevent context poisoning from spreading
Human experts remain essential for understanding product requirements, business context, and unwritten conventions that agents cannot infer from code alone
Agents should be trained to ask for clarification rather than charging ahead, similar to how junior developers need to learn to seek more information before coding
The scattered nature of context across commits, PRs, Slack, and people's heads is a genuine and hard problem that simple tool connections cannot solve

Opposed

The real bottleneck is not context quantity but the ability to focus on recent intent while deprioritizing stale context -- 'context poisoning' is the core problem, not context scarcity
Intelligence remains the actual bottleneck: current frontier models make confidently wrong decisions even on trivial tasks with full context, proving that more context would not help
Benchmarks like ICPC are perfectly suited to LLMs and do not demonstrate real-world coding ability, so the article's premise about superhuman intelligence is flawed
The fundamental architecture of token prediction is the limitation, not context -- no amount of context engineering will make probabilistic generation reliable for programming
Requiring codebases to be refactored and restructured for 'LLM compatibility' defeats the purpose of using the tool to save time and effort
Speed of codebase navigation, not context, is the biggest practical bottleneck for current agents
MCP and context engineering approaches are wasteful since the real gains happen in the first portion of the context window, and expanding it has diminishing or negative returns
Human code review capacity is the true bottleneck -- agents can generate code faster than humans can verify it, creating a responsibility gap