Attention Harness: How to Preserve Human Attention in the Multi‑Agent Era

Background and Motivation

Inspired by the popularity of Codex Pet, the author built a desktop helper called Cliplet that routes user requests to Qoder’s /slide skill to generate PPTs. While developing it, the core challenge shifted from adding another notification to managing how humans monitor multiple agents running asynchronously.

Agents as Stateful Entities

Agents are no longer just stateless models behind an input box; they have state, tasks, waiting points, and can proactively notify users about permission requests, questions, or results. This resembles instant‑messaging (IM) session management, but instead of messages, agents manage execution events such as commands, permissions, evidence, validation failures, and decision forks.

Analogy to IM and Its Limits

The author maps IM mechanisms to multi‑agent concepts:

Online status → Agent running, blocked, waiting for authorization, or failed.

Unread messages → New agent progress or attention‑requiring event.

@me → Agent requests decision, permission, clarification, or acceptance.

Group chat → Collaborative space for multiple agents.

Private chat → Local context between a human and a single agent.

Thread → Long‑term context of a task.

Pin/Star → Critical tasks, evidence, or risks.

Typing indicator → Agent thinking, executing, calling tools, or awaiting tests.

Message alerts → Task status, risk, or completion alerts.

While the IM analogy helps, it stops short because agents manage execution, not just messages. An “unread” agent event may indicate a blocked task, a pending permission, a failed test, or a low‑risk update.

Problem: Reasonable Interrupts Are Still Disruptive

When using tools like Qoder, Codex, Claude Code, or Routa simultaneously, users are overwhelmed by interruptions that all appear reasonable—permission requests, user questions, review requirements, or task completions. However, reasonable does not equal immediate; users may be deep in other work and cannot afford every interruption.

Introducing Attention Harness

Does this agent currently have the right to occupy a human’s attention?

Attention Harness adds a decision layer before an agent event becomes a notification. It classifies events into:

AgentEvent
↓
AttentionHarness
├─ ambient: only change state, no interruption
├─ digest: aggregate for later review
├─ inbox: needs human handling, but not now
├─ interrupt: must interrupt immediately
└─ task handoff surface: present the task context to the human

The system evaluates whether a task is blocked, involves high‑risk or irreversible actions, requires human judgment, lies on a critical path, and whether the human can quickly resume context. Only when all factors indicate “must handle now” does the agent interrupt.

Interaction Shift: From Notification to Task Handoff

Traditional software treats notifications as simple entry points. In the agent era, a notification signals that an execution point needs human involvement—permission, decision, failure correction, result acceptance, or a critical block.

Consequently, the human’s action changes from “reading a notification” to “taking over a task scene.” An example reminder includes detailed context:

Slides generated, preview available.
Task: generate Pet Agent skill presentation
Status: previewable
Why needed: confirm whether to modify further
Recent action: PPT generated, Office WebView launched
Evidence: Canvas artifact / WebView URL / generation log
Possible actions: open preview / continue editing / handle later / archive

The reminder is not just a card; it can appear in notifications, task lists, sidebars, or full task control panels, always providing enough context for immediate handling.

Evolution of Attention Harness

Initially a notification filter, Attention Harness will evolve into a full front‑end scheduling surface as agent count, task length, and execution risk grow:

Step 1: From notification routing to task routing—deciding where a task’s state should flow (ambient, digest, inbox, interrupt).

Step 2: From single pop‑ups to a prioritized queue ordered by risk, blockage, critical path, and context‑switch cost.

Step 3: Into a task‑recovery interface and permission gate, showing target, stage, recent actions, evidence, risk, and actionable options; permission requests will explain why they are needed, associated risks, rollback possibilities, and safer alternatives.

Ultimately, Attention Harness will be a suite of capabilities—interrupt router, state compressor, task‑recovery view, permission gate, evidence entry, and human‑judgment queue—managing the scarcest resource in multi‑agent systems: human attention, judgment, and responsibility boundaries.

Conclusion

While Cliplet and similar agents add cute UI elements, the real discussion is how to design a quiet, recoverable, auditable front‑end scheduling layer that only summons humans when judgment, authorization, acceptance, or correction is truly required.