Real-Time Multiplayer Interfaces

2026-07-05 · source · source ↗ · raw capture · wiki page

Author: Ramon Marc Published: June 8, 2026 Publication: Mold&Yeast (Substack)

Summary

As AI agents become durable — holding state, running on heartbeats — human-agent interaction shifts from one-shot prompting to continuous, real-time dialogue. The interface is no longer a static screen but a "shared protocol" between both parties. Marc argues that "cronjobs are having a comeback" and that once both sides persist, every interface becomes multiplayer — requiring real-time design.

Real-time here means "a fluid dialogue, where both sides continuously listen, respond, and can interrupt" — not just fast turnarounds. Human roles shift from operators at the center to "initiators at the first mile, confirmers at the last, or observers a mile above."

Core Arguments

Interruption Goes Both Ways

Bidirectional interruption is central. Agents cutting into human thought is as important as humans cutting into agent tasks. "If you think your attention is fractured, wait till you unleash the swarm."

The Asymmetry Problem

For agents, reading and responding costs computation. And computation is buyable. For us it costs attention. And attention can only be outsourced, never purchased.

This imbalance means "prioritization and misunderstanding become the things we are actually designing around."

From Screens to Systems

We stop designing fixed moments.

Instead, designers create the "constraints, primitives, and protocols" from which the right interaction ideally emerges in the moment.

Five Principles for the Loop

  1. Mutual readability — a two-way observable loop, not a one-way adaptive surface
  2. Appropriate load — cost proportionate to payoff on both sides; "demand more attention" is not a solution
  3. Interruption as a first-class act — when to interrupt and when not to, in both directions
  4. Derived, not fixed — interfaces generated to match intent and context
  5. Learns with us, does not calcify — grows through a two-way loop, avoids compounding preferences into bias

The Wicked Problem

The fifth principle is the hardest. A shared abstraction layer "can carry quiet bias" and starts cold with no precedent. An adaptive interface can only respond well to situations it has seen before — everything else is a guess. Proposed solution: rehearsal — "small, recurring, Duolingo-like interactions" to surface edge cases before high-risk moments.

Open Questions

Meta Notes