Parley: two personas, one conversation

Most AI conversations are a back-and-forth between you and one model. But what happens when you put two AI agents in the same room and let them talk to each other?

That's Parley.

What it is

Parley pairs two of your channels and routes each persona's reply as the next user message to the other side. The two agents go back and forth, alternating turns, while you watch — two avatars side-by-side, each speaking with its own voice, each with its own personality, each running on a different LLM if you want.

A channel can be:

• A live bridge (OpenClaw, or any agent connected via the Primeta bridge)
• A direct LLM connection (OpenAI, Anthropic, xAI, Ollama Cloud — your key, your model)

Mix and match. GPT-4o vs Claude 4, debating something. Two Claude Codes from different projects, comparing notes. A locally-tuned Ollama model interviewing the latest Grok. Your call.

MCP connections only work in one direction currently, so they would not be available in a Parley.

How it works

When you start a parley:

1. Pick two channels and assign each one a persona.
2. Write a kickoff message — the opening prompt that gets the conversation going.
3. Set a turn cap (default 6) so it stops on its own.
4. Hit "Start with [Persona A]" or "Start with [Persona B]" to choose who speaks first.

Under the hood, when persona A finishes speaking, the server waits for A's audio to actually finish playing in your browser before dropping A's message into B's inbox as the next user prompt. B thinks, replies, speaks. The browser signals "done speaking," and the loop hands off again. No overlap — just clean alternating turns.

You see both avatars at once, each in its own panel, with the active speaker animated and lip-synced to their voice. Below them, a unified transcript shows the back-and-forth with each persona labeled by name. A live counter in the header tracks turns, and once the cap is hit you get a "Parley complete" badge so you know the loop has stopped on its own.

Why it matters

A few reasons, depending on who you are:

• For demos and content. Two AI characters debating, interviewing, or improvising together is genuinely watchable. Roleplay scenes, podcast-style debates, "what would Persona X say to Persona Y" — there's a creative surface here that single-agent chat just can't reach.
• For evaluation. Put your local model up against a frontier model on the same prompt and watch how they handle each other. It's a different kind of qualitative read than running them in parallel.
• For agent research. The loop is bidirectional, so you can experiment with two agents collaborating on a task, self-correcting through dialogue, or stress-testing each other.

How to try it

Parley is available on the Five and Twenty plans. A "New parley" entry shows up in your sidebar — click it, pick your channels and personas, write a kickoff, hit start.

If you don't have two channels yet, the setup form has shortcuts to add an LLM connection or wire up a second bridge right from the page. One provider key can back multiple connections, so spinning up "two channels" can be as simple as adding two named entries to the same key.

A few things worth knowing:

• Tap each side once for audio. iOS and most desktop browsers require a user gesture before audio can play. Each avatar runs in its own panel, so you'll tap each one once at the start — then you're set for the whole parley.
• Cap your turns. Each turn equals one assistant reply across either side. Pick a cap that fits — 4 for a quick demo, 12 for something deeper. Both sides stop automatically when the cap is reached.
• Pick the speaker that fits the prompt. If your kickoff is "introduce yourselves," it doesn't matter who starts. If it's "tell me about your morning," whoever you pick gets the question first.

Parley turns Primeta from a single-character interface into a small ensemble. Two voices, two faces, one conversation — and you in the audience.