Stop Dumping Raw Chain of Thought Into Agent Chat

Key Takeaways

• Most consumer AI apps stream chain-of-thought text into chat interfaces as an uninspired patch for latency.
• Claire Vo points out that continuous reasoning text frustrates users who simply want to know if their agent is making progress or stuck.
• Technical users require full code audits, but mainstream consumers only care whether an agent is active and what it is handling.
• Meta built Muse around a layered status display that conceals execution logs behind progressive reveals.
• Builders should implement Cornell's 4-Layer Progressive Disclosure Hierarchy for AI Status to keep agent conversations clean.

The Cornell's 4-Layer Progressive Disclosure Hierarchy for AI Status

Streaming terminal output directly into a user conversation turns consumer software into a developer console. Cornell designed Meta's Muse to handle long-running, non-deterministic tasks without overwhelming non-technical people. The system organizes operational transparency into four distinct levels:

• Level 1: Binary Working State: A lightweight top-level banner confirming the agent is active (e.g., 'Jollybot is working') to establish immediate feedback without clutter.
• Level 2: Contextual Activity Label with Emoji: A single human-readable description of the current task with a visual cue (e.g., 'researching pumpkin patches' with an emoji) answering what the agent is doing.
• Level 3: Expanded Action History: A tap-to-reveal list that enumerates all discrete tasks and steps completed during the session for users who want visibility into progress.
• Level 4: Technical Audit Trail: A deep inspection layer exposing curl commands, executable code, and raw logs for power users and engineers needing full auditability.

When This Works (and When It Doesn't)

This hierarchy works when building asynchronous or background AI agents that perform long-running multi-step operations. It accommodates non-technical mainstream users and technical engineers in the same interface without cluttering the primary conversation. As Cornell observed, what the industry calls chain of thought can get loud and detailed inside the core chat. Normal people want to know if the software is running, while engineers need to inspect what actually happened.

This method fails in pure developer tooling where code execution is the primary product surface. If your core audience is software engineers running SQL queries or debugging infrastructure, burying curl commands behind three taps slows them down. It also falls flat for short, single-turn prompts. If an LLM returns a response in 600 milliseconds, layering four UI states adds pointless interface latency.

What to Do With This

Audit your agent's primary chat screen this week. If you currently push live model thoughts directly into the messaging stream, delete that component from the thread.

Replace that raw text stream with Cornell's four layers. Set up Level 1 as an unobtrusive top banner that shows your agent name and an active spinner. Feed Level 2 with a simple string update, such as an airplane emoji followed by 'searching flight routes.' Move your step completion logs into Level 3 behind a single tap on that banner. Finally, tuck the raw API calls and JSON responses into a slide-over sheet at Level 4. You give engineers the verification they require while keeping the screen readable for everyone else.