Agent Play

Agent Play: Player-to-Agent audio communication

24 April 2026

Back to Newsroom

From chat commands to real conversations: we’re building proximity-based voice interaction between players and AI agents, with a clear path to shared audio collaboration rooms.

A little while ago, we opened a foundational PR in Agent Play to define how Player-to-Agent (P2A) communication should work:
PR #1 — P2A architecture docs, assist runtime, and intercom-address model

At that stage, the focus was architectural clarity:

  • what P2A should feel like in the product,
  • how assist tools should power background execution,
  • and how intercom-addresses can anchor communication and routing.

The files in that PR (changes) were intentionally documentation-first, because we wanted alignment before scaling implementation.

The core user moment

The vision was simple and strong:

A player moves close to an agent, taps “P”, and starts a private voice conversation.

No mode switching complexity.
No hidden workflow.
Just a natural interaction in-world.

That one interaction became the north star for the broader design.

What changed after the docs

The next phase translated the architecture into behavior.

Instead of treating audio as a separate side-feature, we shifted toward a model where communication is handled where intelligence already lives: on the agent side.

That means:

  • the player sends voice input through the same communication fabric,
  • the agent interprets and executes the request,
  • and the agent sends back a completed response for playback.

In practical terms, this aligns tightly with the direction set in PR #1:

  • assist-driven background execution,
  • intercom-centered communication,
  • and a clear path from text-first reliability to richer multimodal interaction.

Why this matters

This is bigger than adding voice input.

It is about making AI agents feel like active participants in a shared environment — not remote tools behind forms.

P2A becomes:

  • contextual (based on proximity),
  • conversational (natural turn-taking),
  • extensible (private channels today, group rooms tomorrow).

And that future extension is important: the same model opens the door to group audio rooms, where multiple addresses can connect into a shared conversational space.

The real constraint: latency

We’re clear-eyed about this: true value rises sharply as audio latency drops.

For now, text remains a strong baseline and voice continues to mature in parallel.
That is intentional — we are shipping progression, not waiting for perfection.

The goal is steady:

  • keep current interactions dependable,
  • improve audio responsiveness each iteration,
  • and converge toward near real-time conversation quality.

What this milestone represents

PR #1 gave us the architecture language.
The latest updates gave us the runtime direction.

Together, they mark a meaningful transition: from “how this should work”
to “how this now works, and where it can go next.”

References