The Glass Box Engine
An orchestrator that directs, expert agents that execute. This is what turns a request into real action — not just text.
It directs. It executes nothing.
Every request is read by an orchestrator whose only job is to choose among three routes: answer directly, send a scout for a read-only recon before deciding, or break the work into a structured plan handed to experts. It never touches write tools itself — it directs. That decision is semantic, made by the model in any language, never a hard-coded keyword list.
A team of experts, in parallel
The plan is executed by expert agents, each running its own reason → act → observe loop with real tools. Independent tasks run in parallel; each wave automatically inherits the previous one's results. No shared files needed.
create_plan to delegate. No write action by itself.
You set how hard it thinks
From most economical to most rigorous. Higher effort raises the ceilings (depth, agent count, reasoning budget) and adds an automatic reviewer that re-reads every artifact — coverage, consistency, sources. The level guides; it never forces, never changes the guardrails, and stays bounded by an absolute cap (depth ≤ 10).
Direct answer
Fast, little decomposition.
depth 0 · 3 agents · reasoning off
Balanced
Speed and quality.
depth 1 · 8 agents · reasoning low
Fine decomposition
+ quality-review pass.
depth 2 · 16 agents · reasoning medium · review
Critical iteration
Deep work, intermediate results challenged.
depth 3 · 30 agents · reasoning high · review
The "judgment" decisions — should it delegate? is this fact durable? is this signal noise? — are delegated to the model through guidance, never hard-coded. The system works in any language, and the model provider is configurable per deployment.