| name | orchestrate-subagents |
| description | Designs multi-subagent workflows for a ZCode plugin within the 3.5.2 limits — user-scope agents, foreground, parallel-yes/background-no — using Explore for recon and critic panels with on-disk evidence. Use when a plugin needs several agents to cooperate rather than one. |
orchestrate-subagents
Compose ZCode subagents into a reliable workflow. Author each agent with
add-agent; this skill is about how they cooperate given ZCode's real limits.
ZCode 3.5.2 subagent constraints (design around these)
- Agents load from user scope only (
~/.zcode/agents/) — the installer
flattens agents/; a plugin-bundled agent is diagnostic-only until flattened.
- Foreground and parallel are supported; background execution is not (Beta). A
subagent runs to completion in the foreground; several can run in parallel.
- Built-in agents: General-Purpose and Explore (read-only). Prefer
Explore for recon before any write.
- Model per agent:
GLM-5.2, GLM-5-Turbo, or inherit. Invoke with @, or let a
focused description auto-select.
Patterns
- Recon then act. Run
Explore (read-only) to map the code or surface, then a
General-Purpose agent to make changes from that map — writes stay informed and
cheap.
- Parallel fan-out. Split independent work (per-file, per-dimension) across
several foreground agents run in parallel; merge their results yourself.
- Critic panel with evidence. For a quality gate, run N reviewer agents, each
with a distinct lens and a binary yes/no verdict written to an on-disk
evidence file; accept only on a majority. Evidence receipts make the gate
reproducible.
- Single responsibility per agent. Each agent's
description names one job so
auto-selection routes correctly; overlapping descriptions cause misrouting.
Procedure
- Choose the workflow shape (recon→act, fan-out, or critic panel) from the task.
- Author each agent with
add-agent — name, model, focused description, and
read-only vs writable tools (least privilege).
- Keep every agent basename unique across the marketplace (the flatten fails
closed on a collision).
- Validate with
validate-components; behaviorally check with devtest-plugin.
Rules
- Never design for background subagents on 3.5.2 — they run foreground.
- Read-only recon (
Explore) before writes; least-privilege tools per agent.
- Reproducible gates write verdicts to disk, not only to chat.
