Agents

Administration → Agents is the product agent catalog. It lists personality, campaign, and platform agents plus the Ecosystem Agent (in-app operator at /assistant). Agents call APIs and MCP tools. They do not replace Runtime POST /invocations or POST /response.

Ecosystem Agent (/assistant) is the Workbench operator. EcoGentic is customer conversational journeys. Data Agents specialists are a separate analytics product. Do not mix those chats.

Catalog

Rows come from the Workbench agent catalog registry. Coverage is wired, partial, or gap.

Id	Title	Category	Typical tools / paths
`spend-personality`	Spend Personality Agent	Personality	`/public/v1/spend-personality/configs`, `/runs`, drift
`money-personality`	Money Personality Agent	Personality	`POST /api/v1/algorithms/money-personality/process` via `call_endpoint`
`predictive-personality`	Predictive Personality Agent	Personality	Compose spend/money with a Runtime score
`conversational`	Conversational Agent	Journeys	EcoGentic configs — not operator chat
`interaction-science`	Interaction Science Agent	Engagement	Sentimental equilibrium process
`intelligent-sales`	Intelligent Sales Agent	Campaigns	List/trigger campaigns
`real-time-recommender`	Real-Time Recommender Agent	Scoring	Workbench `score_customer` / `send_feedback` on a deployment
`dynamic-experimentation`	Dynamic Experimentation Agent	Dynamic	Dynamic Engagement entities; `explain_algorithms`, `explain_convergence`
`claude-operator`	Ecosystem Agent	Platform	In-process MCP; `/assistant`; HTTP MCP attach
`ecogentic`	EcoGentic	Platform	Customer journeys

Open a row for capabilities, MCP tool names, skills, and a Workbench deep link (workbench_href).

Ecosystem Agent (`/assistant`)

The operator chat is Claude Agent branding (not “Claude Code Agent”). Model traffic always goes through Workbench POST /v1/messages → the default LLM (LLMService.get_config("default")). There is no Anthropic API key in the image.

Consumer	Needs	Does not need
Workbench `/assistant`	Ready Agent SDK + default LLM + in-process MCP	Public API key
Campaign Management	Private `/claude-agent/*` with `tool_profile: campaign_management`	Workbench mutating tools (denied on that profile)
External Claude Code / Agent SDK	`ANTHROPIC_BASE_URL` → Workbench `:8001`; optional dual-auth `POST /mcp`	Bundled `_bundled/claude` in the Workbench image

Readiness: GET /api/v1/claude-agent/status → ready: true when enabled + default LLM + SDK + resolved CLI.

Policy:

Always list_api_catalog first
Use MCP tools only (mcp__ecosystem-workbench__*)
Confirm before mutating (call_endpoint, entity, update_current_form)
When [Current form] is present, explain_entity explains it; update_current_form patches the open editor; the operator Saves
Customer conversations stay in EcoGentic

Private routes:

Method	Path	Purpose
`GET`	`/api/v1/claude-agent/status`	Readiness
`POST`	`/api/v1/claude-agent/…`	Chat turns (operator and CM profiles)
`POST`	`/api/v1/agents/entity/explain`	Read-only explain of an entity / current form
`POST`	`/v1/messages`	LLM proxy for the SDK

How agents reach APIs


Operator /assistant
        │  in-process MCP (no public key)
        ▼
Workbench tools → private /api/v1/*  (and remapped /public/v1 paths)

External Cursor / Claude Code
        │  POST :8001/mcp  (JWT + X-API-Key)
        ▼
Same tool implementations → private API or public gateway

Product notebooks
        │  :8004/public/v1  (JWT + X-API-Key)
        ▼
Curated public catalog only

Scoring a customer from an agent:

Workbench tool score_customer uses a deployment on the Workbench/Runtime path the operator configured.
Production apps should still call Runtime POST /invocations with params as a JSON string.
Feedback is Runtime POST /response (singular) with that row’s uuid and offer_name, or Workbench send_feedback which wraps the same loop for a deployment.

Do not send nested JSON objects in params. Do not call Runtime POST /mcp with a Workbench ewb_public_ key expecting scoring tools — that is a different MCP server.

Campaign Management


WORKBENCH_PRIVATE_API_BASE_URL=http://ecosystem-workbench2:8001/api/v1

CM uses tool_profile: campaign_management (CM MCP pack + a read-only Workbench subset). It must not receive private ecosystem JWTs meant for Grafana; mint a public credential for CM (the CM / ewb_public_… row on API Keys).

Skills

Committed operator skills (Workbench .claude/skills/): workbench-navigator, algorithm-selection, dynamic-experimentation, real-time-recommender, conversational, spend-personality, money-personality, predictive-personality, ontology-data, intelligent-sales, interaction-science, claude-operator, model-convergence.