The Bottleneck Resolution Protocol: How We Keep AI Agents From Getting Permanently Stuck

Every multi-agent AI system eventually hits the same problem: an agent reaches a step it can’t complete without human input. How that moment is handled determines whether your automation compounds or constantly breaks down. We built a formal protocol for it.

The Problem With Informal Escalation

Before the protocol existed, agents would either halt silently (bad—nothing gets done), spam alerts (bad—alert fatigue), or attempt workarounds that caused unintended side effects (worse). There was no shared understanding of when to escalate, how urgently, or what information to include when doing so.

The Four-Level Ladder

The Bottleneck Resolution Protocol (BRP) defines four escalation levels, each with specific triggers and required actions:

• L0 — Try Harder: Agent exhausts at least two independent approaches over 45 minutes before escalating. Most soft blocks resolve here.
• L1 — Pull In Other Agents: Consult a secondary AI system or use a different tool. Documented in the bottleneck case file.
• L2 — Coordinated Sprint: Multiple agents coordinate on the problem with a defined time box.
• L3 — Human Handoff: Prepare a complete handoff document and send an alert. The alert includes: what was being attempted, what was tried, what specifically requires human input, and what the agent will do while waiting.

Hard blocks—anything requiring identity verification, financial authorization, or CAPTCHA—skip to L3 immediately. Soft blocks must exhaust L0 and L1 before escalating.

The Case File System

Every bottleneck gets a case file with a unique ID, priority level, and complete audit trail of what was attempted. A shared alerts.md file lists all open cases. Every agent reads this file at startup—if there’s a human-blocked item relevant to their work, they know about it before they start.

Protected Windows

One underappreciated problem: agents piling up low-priority escalations during times when the human is known to be unavailable (e.g., during work hours at a separate job). The protocol defines protected windows during which only L3-priority escalations can interrupt. Lower-priority cases queue silently until the window opens.

What This Looks Like in Practice

On any given week, there are 2–5 open bottleneck cases. Most are soft blocks that resolve within a day: expired OAuth tokens, rate limit windows, third-party APIs returning unexpected formats. The protocol ensures they’re documented, prioritized, and not forgotten—and that agents route around them productively rather than idling while they wait.

The protocol is the single biggest reason the system stays productive over weeks and months rather than degrading into a pile of broken processes. Governance infrastructure for AI systems is not optional—it’s what makes the difference between a demo and a working system.