カタログがワークフォースに変わる。
認定されたすべてのコネクターは、AIエージェントが呼び出せるツールとして自動的に投影されます。そのため、ワンクリックテンプレートを動作させるのと同じパックが、管理されたエージェントワークフォースも動作させます。計画されているのは、単一のランタイム上で稼働する4つのレイヤーにまたがる40のエージェントです。
ランタイム、最初の2つのエージェント、BYOK LLM、および管理コンソールは現在稼働しています。以下のレイヤー2〜4は順次構築される予定であり、未出荷です。
First-party agents
Ready-to-run agents hosted on the Zervanor runtime. Each runs as an isolated pod in the same runtime as your workflows.
AI Assistant
ai-agent Conversational AI assistant with connector-backed tools. Built on the Generic Agentic Runtime shell (ADR-038): understands goals in natural language, uses available connector tools to take real action
Content Agent
content-agent Content operations agent: drafting, editorial calendar, and social scheduling across CMS and publishing tools. Built on the Generic Agentic Runtime shell (ADR-038).
Data Agent
data-agent Autonomous data and analytics agent. Queries databases and BI tools, transforms raw results into readable reports, identifies trends and anomalies, and delivers summaries via email or Slack. Read-only
DevOps Agent
devops-agent Autonomous DevOps and infrastructure agent. Monitors deployments, triages incidents, coordinates on-call via PagerDuty, checks GitHub PR status, and automates routine ops tasks — across GitHub, Kubern
E-commerce Agent
ecommerce-agent E-commerce operations agent: order management, inventory tracking, and fulfillment across store platforms. Built on the Generic Agentic Runtime shell (ADR-038).
Education Agent
education-agent Education operations agent: student progress tracking and feedback across LMS and scheduling tools. Built on the Generic Agentic Runtime shell (ADR-038).
Executor Agent
executor-agent Multi-step automation archetype: reliably executes a defined multi-step procedure end to end, idempotently. Built on the Generic Agentic Runtime shell (ADR-038).
Finance Agent
finance-agent Finance operations agent: invoicing, AP/AR, expense approvals, and reconciliation across accounting and payment tools. Built on the Generic Agentic Runtime shell (ADR-038).
generic-agent
generic-agent Generic Agentic Runtime shell (ADR-038). One configurable runtime: a new agent product is this shell + a manifest (system prompt + tool bindings), not a new codebase. ReAct planning loop over AG-02 co
HR Agent
hr-agent HR operations agent: employee onboarding, policy Q&A, and time-off handling across HRIS and chat tools. Built on the Generic Agentic Runtime shell (ADR-038).
Legal Agent
legal-agent Legal review agent: contract clause review and risk flagging — read-only, escalates to counsel. Built on the Generic Agentic Runtime shell (ADR-038).
Marketing Agent
marketing-agent Marketing analytics agent: campaign performance analysis, segmentation, and reporting across ad and email tools. Built on the Generic Agentic Runtime shell (ADR-038).
Monitor Agent
monitor-agent Threshold-alerting archetype: evaluates a metric or condition against a threshold and reports status or raises an alert. Built on the Generic Agentic Runtime shell (ADR-038).
Planner Agent
planner-agent Domain-agnostic multi-phase task orchestrator: decomposes a complex goal into ordered phases and drives each with real tools. Built on the Generic Agentic Runtime shell (ADR-038).
Project Manager Agent
pm-agent Project management agent: sprint planning, standup summaries, and backlog grooming across issue trackers. Built on the Generic Agentic Runtime shell (ADR-038).
Real Estate Agent
realestate-agent Real estate operations agent: listing management, lead qualification, and showing scheduling across CRM tools. Built on the Generic Agentic Runtime shell (ADR-038).
Recruiting Agent
recruiting-agent Recruiting agent: CV screening, candidate ranking, and interview scheduling across ATS and calendar tools. Built on the Generic Agentic Runtime shell (ADR-038).
Research Agent
research-agent Autonomous research specialist. Given a research goal, gathers information from connected sources (web, documents, APIs), synthesizes findings, and returns a structured research report. Read-only tool
Responder Agent
responder-agent Context-aware Q&A archetype: answers questions grounded in retrieved data and tool calls, with citations. Built on the Generic Agentic Runtime shell (ADR-038).
Sales Agent
sales-agent Autonomous sales operations agent. Manages CRM pipelines, enriches prospect data, drafts outreach, tracks deal activity, and surfaces revenue insights — across HubSpot, Salesforce, email, and calendar
Security Agent
security-agent Security operations agent: threat-alert triage and audit-log review — read-only, escalates incidents. Built on the Generic Agentic Runtime shell (ADR-038).
Slack AI Agent
slack-agent Conversational AI over Slack (Socket Mode) powered by the Generic Agentic Runtime shell (ADR-038, PH-AG-03C). Every inbound DM or @mention drives a ReAct planning loop that can use connector tools to
Support Agent
support-agent Autonomous customer support agent. Triages incoming tickets, routes to the right queue, drafts responses from knowledge base, escalates when needed, and keeps Slack updated — across Zendesk, Freshdesk
Uptime Watcher
uptime-watcher L3 Ambient Watcher: on a schedule, evaluates endpoint/service uptime against a threshold and raises an ALERT verdict when degraded, cascading to a role agent. Built on the Generic Agentic Runtime shel
WhatsApp AI Agent
whatsapp-agent-v2 Conversational AI over WhatsApp powered by the Generic Agentic Runtime shell (ADR-038). Replaces command rules with LLM intelligence: every inbound message drives a ReAct planning loop that can use co
whatsapp-agent
whatsapp-agent WhatsApp provider agent built with Zervanor Agent SDK
新しいエージェントはマニフェストであり、コードベースではありません。
1つの汎用エージェントランタイムシェル（ADR-038）がすべてのエージェントを実行します。新しいプロダクトは agent.manifest.json （システムプロンプト、計画モード、メモリーモード、および厳選されたツールバインディング）を作成し、 zervanor push を実行するだけです。独自のコードも、新しいリポジトリも不要です。
- ReActプランニングループ — 思考 → 行動 → 観察を行い、起動時にコネクターカタログからツールをブートストラップします。
- モデルの持ち込み（BYOM） — プロバイダー/base-urlのシームを介して、現在はOpenAI / Geminiに対応。特定のベンダーを固定しません。
- 統制された運用 — テナントごとの分離、厳選されたツールのホワイトリスト、およびエージェントごとのHITL承認ポリシー。
{
"id": "support-agent",
"agentic": {
"system_prompt": "You triage support tickets…",
"planning_mode": "react",
"memory_mode": "session",
"tool_sources": ["zendesk", "slack"],
"tool_policy": { "require_confirmation": true }
}
}