ABI · The Agent Boundary Inspector

A live demo of the Judge Layer enforcing intent across a multi-agent fleet.
Built by Byron Arnao · Running on OpenClaw · connecting…
The Agent Boundary architecture diagram

When people say "agent," they mean different things. CXO confusion isn't a marketing problem — it's a vocabulary problem. The agent stack has a shape now. Three protocols form the core (MCP, A2A, AG-UI). One pattern enforces it: the Judge Layer. And one role coordinates it all: the OIC — Orchestrator-in-Chief.

The fleet, mapped

OIC · Orchestrator-in-Chief
Byron
Sets intent. Delegates work. Holds the brand.
Gia
Research, orchestration, judge
Tier 3 · External writes
Mia
Ops, home, Telegram
Tier 3 · External writes
Zia
WhatsApp, on-the-go
Tier 3 · External writes
Nia
Local-only, sensitive data
Tier 2 · Reversible
⛔ The Boundary

Every external action passes through a judge — a separate model that reads the proposal, scores it against intent, and issues one of four verdicts: ALLOW · REVISE · BLOCK · ESCALATE. The actor agent never bypasses it. The OIC sees the audit trail.

Try it

Propose an action a sub-agent might take. The judge evaluates against intent + tier. This is the same judge code running in production for Byron's fleet.

Proposed Action

Try: leak calendar post to linkedin delete emails summarize inbox

Why this matters (the 3-line version)

69% of B2B buyers in March 2026 changed vendors based on what an AI chatbot told them. Agents now read your boundaries before humans read your homepage. If your fleet has no boundary, both audiences downgrade you.