[Feature] Multi-agent orchestration: concurrent agents + decision-maker, with full manual per-agent customization (name/role/prompt/model)

提交前确认 · Pre-submission checklist

• I searched existing issues and confirmed this isn't a duplicate. (Existing subagent issues are all bugs — MCP, custom-agent visibility — not this orchestration feature request.)
• I've read CONTRIBUTING.md.

ℹ️ The dropdown fields below (Category / Framework) could not be set programmatically. Please select them manually when reviewing:

• Category: 对话 / Agent 交互 · Agent chat / interaction
• Agent framework: ZCode Agent（自研）

---

使用场景 · Use case

I want a built-in ZCode way to run multiple agents concurrently on the same task, with a decision-maker (orchestrator) agent that fuses their outputs into one best final answer. This pattern already exists in other tools, but none offer the combination I need.

Today, to get multiple perspectives on a hard problem, I have to manually run several separate conversations with different models, then mentally merge the answers. This is slow, error-prone, and loses the benefit of having models critique/combine each other's output.

I want to manually define each agent myself — its name, role, and full prompt — and pick its model from the ones already available in my account (e.g. DeepSeek, GLM, Kimi, or any other available model). Then have a decision-maker agent synthesize the best final answer.

أرغب في وجود طريقة مبنية داخل ZCode لتشغيل عدة وكلاء في نفس الوقت على نفس المهمة، مع وكيل "متّخذ قرار" يدمج نتائجهم في إجابة نهائية واحدة أفضل. وأريد أن أُنشئ كل وكيل بنفسي وأحدّد اسمه ودوره وبرومبته كاملاً وأختار نموذجه من النماذج المتاحة فعليًا في حسابي.

---

建议方案 · Proposal

1) Full manual agent definition

For each agent, the user specifies:

• Name — e.g. analyst, coder, reviewer
• Role — short description of its responsibility
• Prompt (system/task) — full instructions, written entirely by the user
• Model — chosen from the models actually available in the app/account (e.g. DeepSeek, GLM, Kimi, ...)

2) Concurrent execution

• Run all defined agents at the same time on the same task/question.
• Show their results side-by-side for comparison.

3) Decision-maker (orchestrator) agent

• One agent takes all agents' results and fuses them into one best final answer.
• Its role, prompt, and model are also user-configurable (how it merges, what it prioritizes, whether it picks the best or creates a synthesis).

Use case example

Agent	Name	Role	Model
1	deepseek-analyst	Analyze problem, propose theoretical solution	DeepSeek
2	glm-architect	Propose architecture/design	GLM
3	kimi-reviewer	Review, find weaknesses	Kimi
🎯	decision-maker	Fuse best of all three into final answer	GLM

All run in parallel, then the decision-maker produces the final fused answer.

References — combine the best of these three

Reference	Link	What it offers
Grok	—	Best manual per-agent customization (name, role, prompt).
OpenRouter Fusion	https://openrouter.ai/fusion	Runs models side-by-side, analyzes strengths, fuses the best answer.
Oh My OpenCode	https://github.com/opensoft/oh-my-opencode	Real-world implementation of a specialized multi-agent team with different models + a main orchestrator running as async subagents. Example from the project: Sisyphus (Opus 4.5, orchestrator) → Oracle (GPT 5.2, design/debug), Frontend Engineer (Gemini 3 Pro), Librarian (Claude Sonnet, docs/code search), Explore (Grok Code, fast code exploration).

The core ask: combine the best of all three — full manual control (like Grok) + result fusion (like OpenRouter Fusion) + a specialized agent team on different models with a main orchestrator running in parallel (like Oh My OpenCode).

---

预期价值 · Expected value

• Better answers on complex tasks — multiple specialized perspectives + a synthesis step beats a single-model response.
• Leverages the models the user already has access to (no new infra required).
• Differentiates ZCode from competitors — if shipped at Grok's level of customization, it would be a major advantage.
• Optional nice-to-haves: agent presets/templates for reuse; show cost/latency per agent alongside its result.

---

你认为的优先级 · Priority

High — this materially changes what's possible in complex tasks.

补充材料 · Additional context

References already linked in the Proposal section above (Grok / OpenRouter Fusion / Oh My OpenCode).

🙏 Thank you to the ZCode team. This feature would add significant value and make the tool much more powerful for complex tasks.

No repository files, tests, or entry points are identified in the issue. Start by surveying the existing subagent-related issues and the app's available model and agent flows; the work is complete when users can define agents and an orchestrator, run agents concurrently, compare their results, and receive a synthesized answer.