Live channels expand coding agent communication

Coding agent workflows are moving beyond the old prompt-and-response loop. Recent harnesses such as Claude Code and Codex can let outside events reach an active agent, while agents can also message one another during a task. Live communication falls into delegation from a lead agent to a worker, coordination between teammates, and channels that let an external system steer or notify an open session.

Claude Code channels are built as MCP servers with the experimental claude/channel capability. A channel server can push notifications over stdio into an existing session, with instructions that tell Claude how to interpret the event. Two-way designs can add tools for replies, while plugins can package channel servers with mcpServers entries, data storage under CLAUDE_PLUGIN_DATA, and credential prompts that keep sensitive values out of settings.json.

Security depends on more than labeling events as reports. Channel implementations should verify the actual sender, validate untrusted strings from bots, logs or issue text, reduce payloads to closed fields, and independently authorize consequential actions. Permission relay is optional but risky because anyone able to reply can approve tool calls.

Operational limits remain significant. Custom channels require development flags or administrator allowlists, blocked or non-opted-in sessions may drop events silently, stdout must be reserved for the protocol, and each Claude Code session starts its own channel process. Channel servers should stay on the v1 MCP SDK, because a newer MCP protocol revision (2026-07-28) cannot carry channel messages, and Claude Code 2.1.285 starts rolling out default requests for that revision. Codex does not yet have an equivalent feature beyond ExternalMessage in its Python SDK.