Overview · Extreme Networks · 2023–2026
Intent Engineering for Agent Orchestration
System design and interaction patterns for the orchestration layer of a multi-agent platform: the part that understands what a person wants, combines it with where they are, and routes the work to the right specialized agent while keeping the person in control.
- Disciplines
- System Design · Interaction Design · AI Interaction Design
- Categories
- AI · Multi-Agent Systems · Agent Orchestration
Under NDA
This is a high-level overview. The unreleased product and proprietary details are protected, so the work is described at the level of the problem and approach rather than specifics. I'm glad to walk through it in more depth in conversation.
The problem
A platform of specialized AI agents is only useful if it feels like one coherent product. Something has to understand what a person actually wants, in the context they're working in, and route it to the right agent, without making them learn the system's internal structure or write a perfect prompt.
My role
I worked on the system design and interaction patterns for the orchestration layer, the agent that sits above the specialists. That spanned how intent is captured and scoped, how the system infers context before the person even asks, and how control stays with them as the agents act.
Capturing intent
Three patterns carry the intent layer:
- Structured prompting with guardrails: the interface scaffolds a request, surfacing options, constraints, and boundaries up front, so intent arrives clear and bounded instead of as an open-ended prompt.
- AI auto-complete: the system predicts and completes intent as the person types, lowering the effort to express a goal and steering toward what the agents can actually do.
- Action phrases: shortcut phrases that collapse into reusable tokens, learned like keyboard shortcuts.
Context and control
The system reads two kinds of context at once, where the person is and what they've been doing, then merges them into a best-effort read of intent that pre-selects a starting point without locking anyone in. Guardrails keep the person in control: explicit intent overrides inference, low-confidence routing asks before it acts instead of guessing silently, and recommendations are staged for approval rather than executed. The effect is a system that feels proactive without taking over.