Meta Muse: AI Designed for Human Interaction

Muse employs status indicators to transparently communicate its actions to users, managing expectations during AI processing.

Engineers can access technical logs for specific agent actions within Muse, while typical users receive simplified updates like 'It's working' at the screen's top, ensuring clarity without overwhelming them.

The system's design addresses the challenge of background processing by signaling to the user when the AI is active, which helps manage waiting times.

Proactive messaging capabilities and bubble-based layouts fundamentally change agent design, shifting the interface from a software tool to an exchange between entities.

The agent's ability to proactively message users fundamentally changes how the AI is designed, creating a more dynamic interaction.

Bubble-based message layouts separate multi-time interactions, while visual embodiment at the top of the screen humanizes the experience, making the interface feel more like conversing with a person.

Most AI tools emphasize software-centric primitives like threads, diffs, and connectors, but Muse shifts this focus to user-centric concepts.

Muse elevates 'character,' 'chats,' 'ideas,' and a 'feed' as core entities, reflecting a consumer-focused design approach rather than an enterprise-coding mindset, allowing the system to serve a broad range of user intentions and personal use cases.

A primary design goal for Muse was to create an intuitive experience for non-technical users, making complex AI technology accessible.

The success of the agent, even with non-technical users such as the designer's mother, indicates improved accessibility, achieved by managing complexity through progressive disclosure to prevent user overwhelm.

To maintain a clean chat experience while providing depth, Muse organizes updates into simple, contextual, and technical logs.

The top level provides simple updates, such as 'Jollybot is working,' while the mid-level offers contextual information like 'researching pumpkin patches.'

For users desiring deeper insight, the deep level provides technical transparency, including access to logs and code, ensuring that all users can understand the agent's actions.

Many designers struggle with how much to narrate during AI interactions to make the latency less painful for users.

While chain-of-thought elements often act as temporary solutions for latency, excessive narration can overwhelm users, and insufficient updates can cause anxiety; Muse addresses this by using progressive disclosure and optimizing model speed to maintain a responsive interaction.

When processing long-running tasks, the agent immediately acknowledges requests through reactions or initial messages to prevent users from feeling ignored.

The status display bridges the gap during these tasks, providing continuous updates and maintaining a sense of human-like communication flow.

This approach ensures users remain informed about the agent's progress, even when tasks require extended processing times.

Muse moves beyond a single chat window by employing specialized interfaces, such as 'ideas,' to expose additional capabilities that might not be immediately intuitive.

Suggestions are tailored to the user's conversation history, offering personalized content like golf-related ideas for avid players; the feed surface allows for creative experimentation, serving as a curated display powered by user personalization.

Alex Cornell uses Muse to organize Instagram saves into categorized folders, demonstrating its practical applications.

A feature called 'Jollybot Sky Mall' renders shopping suggestions in the style of 90s catalogs, transforming mundane data into a visually engaging format.

The system monitors specific interests, such as new album releases or golf gear collaborations, providing curated updates.

Muse functions as a research assistant, handling price comparisons and sale tracking, thereby replacing manual search processes.

Instagram content has historically been challenging to process or reason with, but Muse transforms this static data.

Through an Instagram connector, the agent extracts value from personal visual data, helping users generate concepts or descriptive language based on saved visual boards, also leveraging Pinterest for managing visual records.

Alex Cornell tracks his golf rounds and individual shots using wearable devices, feeding this data into Muse.

Muse provides a 'round recap' that offers deeper context than a standard paper scorecard, identifying patterns on specific holes and suggesting adjustments to improve scores.

The agent converts raw, monolithic data into visual graphs, telling a clearer story of performance, and practice sessions are analyzed using machine data to provide detailed feedback on swing mechanics.

This data-driven approach allows for continuous athletic improvement by turning simple numbers into actionable insights.

Alex Cornell utilizes the agent's tracking section to monitor long-term goals, such as his aspiration to achieve a single-digit golf handicap.

Side chats are offered for scoping specific tasks or organization, aligning with the philosophy of treating Muse as an entity rather than an app, similar to messaging a person.

Alex Cornell considers current threading an underexplored area within AI interactions.

He believes significant work remains to help users differentiate between main and side threads effectively.

The ultimate goal is to evolve the interface in preparation for the upcoming 'agentic era,' enhancing how users interact with AI agents.

Natural language requests allow users to generate custom dashboards, such as the chess analysis tool Alex created to turn raw data into actionable insights.

For example, Alex Cornell created his chess dashboard by simply asking for specific analysis and trends over time, allowing the agentic approach to convert raw data from various sources into actionable insights.

Designers must shift from controlling every screen element to focusing on the underlying architecture when working with non-deterministic AI systems.

AI experiences are highly personal, leading to significant variations in the interface for each user, requiring designers to embrace 'non-deterministic' outcomes instead of forcing rigid, pixel-perfect layouts.

This shift towards prioritizing the system's foundation over individual components is described as a liberating approach to design.

Reflecting on real-life communication challenges, Alex Cornell notes that poor instructions often drive user frustration, so he maintains a polite tone even when agents like Marlo or Jolly Bud are unhelpful.

Some users employ 'forceful' prompting to initiate a system reset when the AI deviates, though the creator notes that user frustration often stems from poor instructions, mirroring real-life communication challenges.

Feedback submitted directly within the app is reviewed by the development team, and those seeking deeper engagement can follow Alex Cornell on Threads for design reasoning and behind-the-scenes content.