Walk into any engineering leadership meeting today, and someone will question whether AI-generated code is secure or whether agents can be trusted in production. These are valid concerns, but they don’t determine whether your team ships faster. The bottleneck is context: the gap between what engineers carry in their heads and what AI can understand or communicate.
Companies that solve the context problem will move faster. Their tools will make fewer mistakes that require human correction, while teams that ignore it accumulate technical debt in the form of code that developers can’t fully explain. Security and quality matter, but they’re largely addressable at the technical layer; the real constraint is transferring engineers’ tacit knowledge into systems.
Code quality tools are ready, but context isn’t
By 2025, AI code review had truly arrived, and static application security testing (SAST) tools were already catching the obvious issues. Today, most companies run one or more AI reviewers on every change, and false positives are low enough that these tools have earned their keep. The mechanics just work. Claude Code and similar tools showed in 2025 that AI can write substantial, multifile changes that compile and run.
What doesn’t work is the handoff. An engineer spends weeks absorbing not just the technical architecture but also the unwritten rules that govern a codebase: when to prioritize performance over readability, which abstractions the team actually maintains, and how defensive to be about edge cases. When an AI agent writes or reviews code, it operates without that accumulated knowledge. You can feed it documentation, but documentation is always incomplete; it records what someone thought to write down, not the dozens of micro-decisions that shaped the current system.
The two-way problem
Getting context into AI tools requires deliberate effort that most teams haven’t systematized. Engineers need to translate their implicit knowledge into something an agent can parse. Some companies are experimenting with detailed architecture documents that live in the repo specifically for AI consumption, while others are building specialized prompts that encode stylistic preferences. But these are stopgaps. The UX for context handoff remains clunky, and the tooling barely exists.