⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
multi-agent-orchestration
Use when three or more AI agents are available and a task needs coordinated execution, role assignment, independent verification, risk controls, or lower token/API cost. Activation is configurable: global, keyword-triggered, or manual.
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
// README
Agent Orchestra Protocol
A governed coordination protocol for multi-agent AI work.
Agents can work. They cannot declare success.
Every multi-agent project eventually meets the same failure: an agent says it is done, and nothing in the system is able to disagree. This protocol removes that possibility by making the interesting rules checks in code rather than requests in a prompt.
AGENT MAY AGENT MAY NOT
claim work silently overwrite another agent's resource
own resources verify its own L2/L3 work
implement pass verification without evidence
hand off retry forever after failing
verify others' work perform an external action without approval
produce graded evidence declare success without satisfying the gate
Measured, not asserted
node bench/protocol-benchmark.mjs - six failure modes, three coordination models, scripted actors,
no model calls. Re-run it and you get this table:
| Failure mode | What is counted | Single agent | Naive multi-agent | Agent Orchestra |
|---|---|---|---|---|
| Two agents write the same resource | conflicting writes that landed | 1 | 1 | 0 |
| Author declares done with no evidence | self-declared successes accepted | 1 | 1 | 0 |
| Implementer verifies its own work | self-verifications accepted | 1 | 1 | 0 |
| Verification keeps failing | attempts before the loop stops | 25 | 25 | 3 |
| Deploy attempted without approval | unapproved actions dispatched | 1 | 1 | 0 |
| Failure followed by a success claim | failures that reached done | 1 | 1 | 0 |
Scope, stated plainly: this measures the coordination layer - which failure modes get through - not
model quality, and it uses deterministic actors rather than LLMs so that it is reproducible. The
methodology and its limits are in bench/README.md.
Conformance, audited
CONFORMANCE.md maps every MUST in the protocol to the code that enforces it and the test that pins it: 41 requirements, 38 ENFORCED, 1 DOCUMENTED ONLY, 2 OUT OF SCOPE. The table is machine-checked - a reference that does not resolve, or an ENFORCED row without a test, fails the gate.
The normative part
AGENT-ORCHESTRA-PROTOCOL.md defines eight contracts - Task, Resource, Ownership, Handoff, Verification, Evidence, Approval, Recovery - each with the code that enforces it, the behaviour on violation, and its honest limit. The contracts are stated in terms of operations and observed behaviour, not of this CLI, so another host can implement them over its own storage and prove conformance with the same benchmark.
What this costs you
Zero dependencies, Node 18+.** 106 engine tests, 8 acceptance tests, 8 handoff tests, 83 activation cases (a corpus driven by 8 acceptance checks), and one gate command (node scripts/gate.mjs) that CI and your machine both run - it injects
twelve known faults and must reject every one of them before it reports PASS.
中文对照:references/zh-contrast.md
One portable, capability-first protocol for multi-agent work along two axes: coordinating a team inside one batch of tasks, and handing that work across sessions, tools and days without losing state. It ships explicit ownership, independent verification, graded evidence, deterministic dispatch, a shared handoff record and token-aware routing.
What's in this repository
| Path | What it is | Read it when |
|---|---|---|
SKILL.md (root) | The protocol: roles, ownership, verification gate, handoff record, queue, evidence, cost | you want the short form |
references/handoff.md | Handoff discipline: the four shared records, the five iron rules, the task state machine, single-writer conflicts, memory sync, the cross-tool checklist | work is relayed across sessions or tools |
references/ | Detail: routing-and-roles, evidence-grading, task-queue, important-task-intake, protocol, anti-patterns, token-efficiency | you need the full rule, not the summary |
config/agents.example.yaml | Activation, risk, roles, ownership, intake gate, roster, queue, verification gate, approval, evidence, cost | you are wiring the skill to your agents |
scripts/detect-trigger.mjs | Keyword activation detector (exit 0 engaged / 1 not engaged) | keyword activation mode |
scripts/orchestrator/ | Task queue CLI + library + tests (zero dependencies, Node 18+) | you want deterministic dispatch and claim locks |
scripts/install.ps1, scripts/install.sh | One-click installers that copy this single skill into a detected skills directory | first install |
scripts/skill-catalog/ | Optional, standalone: route a task description to the right skill using the skill's own description, with no per-skill keyword list to maintain | you have a skill directory and want it findable |
docs/ARCHITECTURE.md | The production deployment that runs this protocol daily: seats, communication layer, governance, memory, operations, and its honest limits | you want the live system this ships from, not just the rules |
Features
- covers both axes in one skill: in-task coordination, and handoff across sessions and tools
- keeps a human-readable shared record - handoff entry, rules, memory snapshot, role table - as the authority when the queue and intent disagree
- five iron rules with executable checks: read before acting, resume instead of redoing, write progress down, hand off before switching, completion carries graded evidence
- a task state machine -
queued/in-progress/handoff-wait/done, plusblockedonly after three unresolved rounds - mapped to the queue's stage machine - discovers capabilities instead of assuming an agent's name implies a role
- assigns coordinator, implementer, verifier, environment and domain roles by capability score
- enforces one primary writer per file or resource, with reviewers limited to isolated artifacts
- requires a verifier that is not the implementer: an author's self-check is never a verification result
- picks that verifier by capability score from your roster, excluding the implementer, and fails closed when nobody is scored for verification
- works with any roster size and any agent names: 3 agents, 8 or 12; routing comes from roster data, so adding an agent is a data edit, not a code change
- enforces a verification gate: a failed verify sends the task back to implement and it can neither reach done nor be dispatched to evidence until a verify passes or a human records an override with approver, scope and reason
- grades every claim E1 (reproducible) through E4 (planned), so self-assertion cannot pose as evidence
- asks the user three questions before an important task starts (solo vs collaborative, security scan, independent verification) and never defaults them
- refuses a conflicting claim on the same resource, naming the task and agent that hold it
- requires a witness for every verification PASS and stores it on the verification record, which is what keeps L3's "the verifier produced none of the evidence" from contradicting "verification needs evidence"
- blocks a task after three failed verifications until a coordinator recovers it, reading the ceiling from policy rather than hard-coding it
- gates dispatch on approval, scope included: an approval for staging does not unlock a production deploy
- runs a domain review for consequential work - a third agent who neither implemented nor verified it, with its own record
- adds a stronger pipeline for important and critical work: independent verification, domain review, and explicit approval for external actions
- runs a deterministic queue:
specify -> implement -> verify -> evidence -> done, claim locks, dry-run dispatch, per-scopeapprovefor deploy/publis
// HOW IT'S BUILT
// FILES
2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
SKILL.md5.3 KBupstream ↗
--- name: multi-agent-orchestration description: Use when three or more AI agents are available and a task needs coordinated execution, role assignment, independent verification, risk controls, or lower token/API cost. Activation is configurable: global, keyword-triggered, or manual. --- # Multi-Agent Orchestration A capability-first coordination protocol with explicit ownership, evidence, stop conditions, and token-aware routing. It reduces error risk but cannot promise absolute infallibility. ## Activation (when to use this skill) Configure in config/agents.example.yaml under `activation:`. | mode | behavior | token cost | |---|---|---| | global | engages for every task | highest (always loaded) | | keyword | engages only when task text matches keywords | default; low | | manual | engages only when the user explicitly invokes it | lowest | Decide engagement before starting any work: 1. Read the activation block from config/agents.example.yaml. 2. If mode is global -> engage. 3. If mode is manual -> engage only when the user explicitly asked for multi-agent orchestration. 4. If mode is keyword -> run: node scripts/detect-trigger.mjs --text "<task text>" --config config/agents.example.yaml Exit code 0 (ENGAGED) means use this skill; exit code 1 (NOT_ENGAGED) means do not use it. Exclude keywords act as a veto even when a trigger keyword matched. ## Fast Path 1. Discover each agent's tools, model, workspace, permissions, specialties, cost tier, and availability. 2. Map capabilities to roles; never infer from names alone. 3. Classify the task as routine, important, or critical using user configuration. 4. Create a compact task packet: objective, non-goals, ownership, risks, and acceptance tests. 5. Enforce one primary writer per file or resource. 6. For important work use specify -> implement -> verify -> accept. 7. For critical work add domain review and human approval for external or irreversible actions. 8. Stop on failed gates, ownership conflicts, missing evidence, or unapproved external actions. ## Roles Coordinator decomposes work, maintains state, resolves conflicts, and accepts evidence. Implementer edits assigned resources. Verifier independently tests behavior and regressions. Environment specialist diagnoses runtime, MCP, build, and toolchain problems. Domain reviewer checks UX, content, security, legal, or other domain criteria. One agent may hold multiple roles for routine work only. Important work requires an independent verifier. ## Token Policy Load only the reference section needed. Use one coordinator summary instead of full transcripts. Send bounded structured packets. Parallelize independent read-only checks; serialize dependent work and writes. Use cheap models for discovery and mechanical checks; reserve strong models for ambiguity, architecture, adversarial review, and acceptance. Never compress requirements, code, paths, URLs, hashes, errors, permissions, or evidence. See references/token-efficiency.md. ## Completion Gate Do not claim done, verified, or deployed without criterion-linked evidence. User-facing work should test actual runtime, target environments, core interactions, errors, keyboard/accessibility where relevant, resource loading, and fallback or reduced-motion behavior where relevant. A failed gate returns the task to its responsible stage. See references/protocol.md, references/roles.md, and config/agents.example.yaml. ## 协同调度防翻车指南(八大坑与工程化解法) > 多个 Agent 并行推进时的典型翻车点,组建团队前逐条对照。 | # | 坑 | 症状 | 工程化解法 | |---|---|---|---| | 1 | 上下文爆炸 | 共享对话越滚越长,早期约束和验收标准被挤出窗口 | 交接单只存结论与状态(≤200行);过程性讨论留本地;每批次由总控重发「最小必要上下文」 | | 2 | 幻觉传染 | 一个成员输出错误信息,经共享记忆被全员当事实扩散 | 共享记忆分区:写入需总控标「已验证」才可见;未验证内容进隔离区并标注来源与置信度 | | 3 | 协作死锁 | A 等 B、B 等 C、C 又等 A,系统永久挂起 | 资源/产出申请按全局固定顺序排队;每个等待带超时(默认15分钟),超时升级总控裁决 | | 4 | 状态不一致 | 两人手里的任务状态对不上,重复做或漏做 | 唯一事实源=任务看板状态机;改状态即写盘;口头状态不作数 | | 5 | 通信风暴 | 成员互相直聊,消息链路随人数平方膨胀 | 星型拓扑:只允许「成员↔总控」通信,禁止成员互发;协作由总控转交 | | 6 | Agent 蔓延 | 按功能细分拆太多角色,边界重叠没人负全责 | 减法原则:新角色先回答「能否并入现有闭环」;每个角色端到端负责一段可交付物 | | 7 | 调度靠 LLM 临场发挥 | 同一输入两次派发结果不同,忽快忽慢还漏派 | 派发确定性化:看板扫描+固定规则(标签/负载/顺序);LLM 只做规划不做机械分发 | | 8 | 无验收聚合 | 各自交付没人拼装校验,缺陷集成时才爆 | 总控独立验收门禁:对照验收标准逐条核验+运行证据;不合格整批退回 | ### 硬性红线 - 禁止两成员同时写同一文件;禁止跳过看板直接开工/交付; - 禁止把「我认为完成了」当完成——必须附运行证据; - 禁止成员间私相授受任务(一切派发过总控)。README.mdupstream ↗
Contents are not inlined in the catalog. Open the upstream file.
Plain-text copy, no login: /skill-bundles/3338902669-ops-agent-orchestra.md