Context engineering

Context engineering treats the prompt window as a budget to be allocated, not a bucket to be filled. Every token competes; the planner decides what earns a slot.

Navigation over retrieval

Rather than dumping the top-k passages, we expose a navigable structure — sections summarised, leaves embedded — and let the agent walk it. Embeddings act as navigation hints; the real extraction happens at the observer.

1
2
3
4

budget = Budget(tokens=8_000)
for node in planner.walk(doc):
    if budget.fits(node):
        ctx.add(node)