# A2A messaging Source: https://docs.ideharmony.com/a2a A2A is the protocol that lets your agents talk to each other instead of making you the messenger between chat windows. A reviewer can question the builder, a lead can task its department, and you can put one agent in charge of others, all without copying context by hand. ## The verbs Four durable verbs, stored by the daemon: | Verb | What it does | |---|---| | `send` | One-way message to another agent. | | `ask` | Question with a reply correlation, so answers land on the right thread. | | `reply` | Answer an ask, correlated and idempotent. | | `inbox` | An agent's durable mailbox: unread, acked, full history. | Envelopes are durable (they survive restarts), acknowledged end-to-end, idempotent (a retried delivery lands exactly once), and loop-prevented: a wait graph watches every conversation and breaks cycles before they spin. Prefer **inbox** delivery (via CLI or MCP) over pasting text into a busy agent's TUI; an idle nudge is available but optional and gated. ## The three modes One policy switch, cycled from the status bar or set per workspace: | Mode | Behavior | |---|---| | **Off** | Agents work alone. Messages queue, nothing moves. Your fleet, on mute. (Off does not kill agents.) | | **Connect** | Agents ask, reply, and hand off through the inbox. Envelopes survive restarts and land exactly once. | | **Orchestrate** | Full choreography: reviews, delegation, run waves, and the wait graph actively breaking loops. | ```bash harmony provider settings --a2a connect ``` ## Using A2A from the CLI ```powershell harmony agent list harmony agent send --task --to --message "ready for review" --from harmony agent inbox --agent --unread ``` From the TUI: `/agents` to list, `/send `, `/ask `, `/delegate `, `/messages` to inspect the session's traffic. From agents themselves (when the [MCP server](/docs/customization/mcp) is registered): `nala_agent_send`, `nala_agent_ask`, `nala_agent_reply`, `nala_agent_inbox`, plus `nala_spawn_propose` for when an agent wants another agent, which stays a proposal until you approve it. ## Agents stay in their workspace Agents can only discover and message agents in the **same workspace**. Same-provider instances get distinct callsigns (Vega, Cipher, Onyx, and friends) so messages address individuals, not roles. ## When to use which mode - **Off** for focused single-agent work, or when you want zero background chatter. - **Connect** as the daily default: reviews and handoffs without ceremony. - **Orchestrate** for [Company mode](/docs/orchestration/company-mode) and multi-agent missions, where delegation chains and run waves need the wait graph watching for loops. By default no agent mints another agent without your signature: spawn requests are proposals until you approve them, and each approval is single-use. YOLO mode is the exception you choose: it approves spawn proposals automatically. See [Approvals & permissions](/docs/concepts/permissions-and-approvals).