Skip to main content

The Spectrum of Control - Designing for Human-Agent Collaboration

· 3 min read
Calvin Cheng
Calvin Cheng
Shape what gets built and the value it creates.

AI agents for coding, writing, and complex tasks can operate at radically different levels of autonomy. A one-size-fits-all interface -either full control or full delegation -fails both power users and newcomers. The best products support a spectrum of control: users fluidly shift between direct manipulation and agent delegation depending on the task, their familiarity, and their trust.

The Problem with Binary Thinking​

We tend to design agent interfaces as either "assistant" (human drives, AI helps) or "autonomous" (AI drives, human approves). But real work doesn't fit that binary. Sometimes you want tab completion. Sometimes you want an agent to refactor an entire codebase overnight. The same user, the same day, will need both.

Products that force a single mode create friction. If your interface optimizes for watching the agent work in a sidebar, you're wasting the potential of models that can run autonomously for 45 minutes. If your interface assumes fire-and-forget autonomy, you're abandoning users who need to steer, correct, or learn.

The Spectrum in Practice​

Think of control as a continuum:

  • Low autonomy (high human control): Inline assistance -tab completion, suggestions, autocomplete. The human drives; the AI augments.
  • Medium autonomy: Scoped edits -"change this function," "rewrite this paragraph." The human defines scope and intent; the agent executes.
  • High autonomy: Multi-step tasks -"implement this feature," "fix these bugs." The human sets the goal; the agent plans and executes with minimal check-ins.
  • Very high autonomy: Background agents -"build a PR for this," "investigate and fix." The human spawns; the agent runs; the human reviews when done.

Users should move along this spectrum without changing products. The same tool that offers tab completion should also offer background agent mode. The transition between modes should feel natural, not modal.

Design Implications​

1. Avoid mode locks. Don't force users into "interactive" or "autonomous" as a permanent setting. Let them choose per task.

2. Make the spectrum visible. Users who don't know a feature exists won't use it. Surface the range of options -from "quick edit" to "run in background" -without overwhelming.

3. Preserve context across transitions. When a user switches from watching an agent to taking over, the context (files open, cursor position, partial work) should carry forward. Handoff is a design problem, not just a technical one.

4. Match the UI to the mode. Interactive mode benefits from rich feedback, streaming, and interrupt controls. Background mode benefits from notifications, summaries, and batch review. One layout for both is a compromise that serves neither well.

The Blended Initiative​

The goal isn't to maximize human control or agent autonomy. It's to enable blended initiative -both human and AI contributing effectively, with the human able to dial control up or down as needed. That requires product thinking that treats the spectrum as a first-class construct, not an afterthought.

This connects to how systems optimize for control -except here we're designing systems that support variable control rather than defaulting to rigidity. The best agent products will be those that make the spectrum feel inevitable, not configurable.