We process code through successive stages: symbol level, module synthesis, and codebase-wide distillation. Multiple passes allow us to be exhaustive and refine iteratively across levels of abstraction.

Keeping irrelevant, hallucination-inducing content out is as important as getting signal in. Our compiled output separates concerns across concepts and levels of abstraction so agents fetch exactly what they need without polluting their context.

Not all codebase understanding should be pre-computed, and not all of it can wait until runtime. We focus compile time on content agents are guaranteed to need but cannot construct ad hoc.

Good context is wasted if agents don't know how and when to use it. We provide agent-agnostic guidance and orchestration via conventions like AGENTS.md, steering in MCP descriptions, and abstractions like skills.

We can't pre-compute the perfect context package for every possible task. We bundle guided, iterative use of our primitives into higher-level agentic tools that run in isolation to deliver high SNR context packages for a target task.

Users and agents can produce and register their own content into Driver's context layer. Content tied to your codebase stays automatically up-to-date as code evolves, and your team can shape the layer to how you work.