Session Chats: Multi-Agent Conversations

Session Chats allow you to converse directly with multiple AI coding agents in a unified conversation thread. You can brainstorm ideas, ask technical questions, or let agents debate implementation approaches before launching a full loop.

Collaboration Modes

Inside the session chat header, turn controls (#collabTurnControls) allow you to select one of three collaboration modes:

1. Standard Mode

Behavior: Only agents specifically addressed in your prompt will reply.
Use Case: Quick questions to a single agent or direct back-and-forth interactions.
Handoffs: Optional. If enabled, an agent can hand off its answer to a colleague using an @mention.

2. Discussion Mode

Behavior: Agents converse with each other across a set number of rounds until they reach consensus or have nothing new to add.
Rounds Selector: Choose from 1 to 5 rounds using the dropdown (#collabDiscussionRounds: 1 round, 2 rounds, 3 rounds, 4 rounds, 5 rounds).
Use Case: Comparing technical trade-offs (such as choosing between database schemas or API architectures).

3. Council Mode

Behavior: Every active agent provides an independent answer to your prompt in parallel. Once all answers arrive, a designated moderator agent synthesizes them into a single, unified summary.
Council Moderator: Select the moderator agent using #collabCouncilModerator (for example, Claude or Antigravity).
Use Case: Getting diverse architectural opinions followed by an unbiased executive summary.

Addressing Agents with @Mentions

You can direct messages to specific agents or address the entire team using @mentions anywhere in your prompt:

@claude: Addresses Claude Code
@gemini: Addresses Gemini CLI or Gemini Web
@antigravity: Addresses Google Antigravity
@hermes: Addresses Hermes
@opencode: Addresses OpenCode
@all: Directs the prompt to all active agents in the session

When an agent replies, it can also use @mentions to request input from a peer agent.

Conversation Handoffs

Under the chat controls, locate the checkbox labeled: "Agents can hand off the conversation" (#collabAgentHandoff).

When Enabled: If Claude determines that a task requires fast code execution or web browsing, it can tag @gemini or @antigravity in its response. Zenborg automatically invokes the tagged agent in the next turn.
When Disabled: An agent’s response ends the current turn, waiting for user input.

Relay Guardrails

To prevent runaway conversation loops between autonomous agents, Zenborg enforces a hard limit:

Maximum Relay Turns: MAX_TURNS = 10. A single conversation chain will never exceed 10 consecutive automated agent turns before stopping and prompting for human confirmation.

Switching Between Chat and Goal Mode

Session chats are designed to flow naturally into autonomous coding:

Chat vs Goal Mode: Toggle between exploratory conversation and structured goal specification.
Open in Loops: Once an architectural discussion reaches a clear plan, click the "Open in Loops" button to transfer the conversation context directly into an autonomous Agent Loop.

Next Steps

Learn how to structure projects in Projects and Goals.
Explore browser-based models in Browser Workers.