Expert agent (agent mode)

An AI Agent node in Consultable expert mode does not process the document. It answers another agent's questions within the pipeline. The agent that consults it gets a consulter_<expert name>(question) tool, which it calls when it needs to, like an MCP tool or a skill.

Examples:

• a form-filling agent asks an expert wired to the contracts base for a customer's member number;
• a support chat agent consults a legal expert before answering;
• an annotation agent checks a regulatory reference with an expert that holds the standards knowledge base.

Wiring

Agent (caller) ◀── experts ── Agent (Consultable expert) ◀── RAG / MCP / skills

Connect the expert's consult handle (top) to the experts handle (bottom) of the calling agent. The caller can be in solo, chat, form filling or expert mode. One expert can serve several callers, and one caller can have several experts.

The expert keeps its own attachments: knowledge base (RAG), MCP connectors, skills, HNS packages, and its own experts.

An expert does not go inside a Parallel container and does not plug into a supervised agent: those agents already run in parallel, and a consultation would add one more execution level.

Configuration

• Agent: the Designer agent that answers. Its role, objective and constraints define its field.
• AI model: optional override of the agent's model.
• When to consult it: what the calling agent reads to decide whether to consult the expert, for example « Knows each customer's member number and contracts ». Empty, it is the agent's description, or else its objective. The more precise it is, the better the caller knows when to use it.
• Consultations per run (5 by default, 50 at most): maximum number of questions asked to this expert during one job, all callers combined. Beyond it, the tool answers « consultation limit reached » and the caller goes on without it.
• Use support documents: unchecked by default. Checked, the expert sees the job's support documents.
• Execution loop: the expert's budgets when it uses its own tools (turns, duration, tokens).

How a consultation goes

1. The calling agent asks a self-contained question. The expert sees neither the caller's document nor its conversation: the question must carry the useful context.
2. The expert answers like a chat turn without history, with its role, its knowledge-base excerpts (queried with the question), its tools and its experts.
3. Its answer goes back to the caller as a tool result, cut at 8000 characters. An expert error also comes back as an answer (« Erreur expert… »): the caller can do without it, and the job goes on.
4. The same question asked twice in a row is served from the caller's cache, without a new call to the expert.

Experts of experts

An expert can itself consult experts, 3 levels at most below the calling agent. The pipeline is refused at save time when the chain is deeper, or when the consultations form a cycle (A consults B, which consults A).

Costs and trace

• Each consultation is a full turn of the expert (several model calls if it uses tools). It is billed to the expert agent: its tokens show under its name in Finances.
• The caller's loop budgets (tokens) do not count its experts' tokens. The Consultations per run cap and the expert's own budgets bound that cost.
• In the test panel, the call shows among the caller's tools (« Expert · name »). Each expert consulted has its own entry, with its questions, answers, level and the tools it used.
• An expert does not move the progress bar: it works during the turn of the agent that consults it.

Cancellation

Stopping a job also stops the consultations in progress: the expert checks for cancellation before answering and between the turns of its own tools.