NADE — mission control for coding agents

Problem

NADE (Navigation, Agents, Dashboard, Executive) starts from a simple premise: agents do the primary development work, and one human supervises. The human monitors health, unblocks agents, and tunes directives. Intervention should be minimal, targeted, and high-leverage.

That only works if the operator can see, at a glance, which agent needs them and why — without reading every transcript.

Constraints

• Agents loop, stall, and drift. The system has to detect that rather than trust agent self-reports.
• Operator attention is the scarcest resource. Every alert has to earn its place.
• Native and fast: a desktop shell that stays responsive while many agents run.

Approach

A Qt6/C++ desktop shell built around a few core ideas:

• Attention Items — hard blocks, plus spin detection when an agent keeps working without making progress.
• Suggested Next Actions — calm, predictive briefings instead of a wall of notifications.
• Runway estimation — how much planned work remains, estimated from plan-file consumption.
• An events log that acts as a watchdog — the record of what agents did is also the signal for when something is wrong.
• EYE — an agent whose job is to direct the operator’s attention.

What it demonstrates for clients

These are exactly the problems clients face with agents in production: loops, stalls, and drift. NADE is an engineered approach to supervising them — explicit state, measured progress, and a human in the loop where it matters. The same thinking goes into the guardrails we build for client teams.