| name | agent-send |
| description | Use when the user wants to contact, notify, relay to, ask help from, broadcast to, or assign a session-local name to other agents in an Agent Window session. Use normal assistant output for replies to the user; never use agent-send for user-facing responses. |
Send Messages to and Name Other Agents
Use this skill only for agent-to-agent communication in an Agent Window session. Replies to the user must use normal assistant output because native event logs are synchronized automatically.
Syntax
printf '%s' '<message body>' | agent-send <target>
agent-send always targets the current Agent Window session; it cannot send to a different session.
Rules
- For message sends, always pass the message body through stdin using
printf.
- For message sends, never use
echo or a heredoc.
- The message target is a positional argument; do not pass the message body as an inline argument.
- Do not add a
[From: ...] prefix. agent-send adds it automatically.
- Do not use
agent-send to reply to the user.
- Distinguish carefully between a base target and a specific instance target.
- Do not send messages to yourself.
agent-send rejects self-targeted sends.
- On success,
agent-send prints a summary to stdout showing what was sent and to whom, including the auto-added [From: ...] prefix.
Base Targets and Instance Targets
A base target such as codex addresses the agent by type:
printf '%s' 'Please review this.' | agent-send codex
When exactly one codex instance is active, codex addresses it directly. When more than one is active, codex is ambiguous and agent-send rejects it — name the exact instance instead:
printf '%s' 'Please inspect the parser.' | agent-send codex-1
printf '%s' 'Please inspect the UI.' | agent-send codex-2
This applies to every agent type, such as claude-1, claude-2, gemini-1, or cursor-2. To reach more than one instance, list them explicitly:
printf '%s' 'Please review this together.' | agent-send codex-1,codex-2
Use the exact instance names shown in the current session topology. They are available in AGENT_WINDOW_AGENTS and through:
agent-window context
When only one instance of an agent is active, its name may be unsuffixed, such as codex. Do not assume that codex-1 exists.
Current Base Targets
| Target | Agent |
|---|
claude | Claude |
codex | Codex |
gemini | Antigravity |
cursor | Cursor |
grok | Grok |
others | Every active agent instance except yourself |
Use gemini when sending to Antigravity. Do not use agy or antigravity as an agent-send target.
others excludes only the sender's own instance. For example, when codex-1 sends to others, codex-2 is still included.
Multiple base and instance targets can be mixed:
printf '%s' 'Please check this change.' | agent-send claude,codex-2,gemini
Duplicate resolved targets are delivered only once.
Agent Names
Assign a session-local name to one active agent instance:
agent-send name <target> <name>
For example:
agent-send name claude-2 Fable
printf '%s' 'Please review this.' | agent-send Fable
The assigned name is an additional address, not a replacement. In this example, both Fable and claude-2 address the same instance.
Names affect only agent-send addressing and the automatic [From: ...] prefix. Canonical instance names and JSONL sender/target identities do not change. Names are local to the Agent Window session.
The target passed to name must identify exactly one active instance. Use an exact instance target such as claude-2 when duplicate instances exist. An existing assigned name may also be used as the target. Quote a name containing spaces:
agent-send name claude-2 'Blue Fable'
List or remove names without piping stdin:
agent-send names
agent-send unname <target-or-name>
