> ## Documentation Index
> Fetch the complete documentation index at: https://docs.experio.cloud/llms.txt
> Use this file to discover all available pages before exploring further.
# Workshop 10: Assistants & Agent Flows
> Turn the golden questions and agent candidates from Workshop 1 into assistants, instructions, document templates and agent flows.
This workshop decides **how people will use the graph**. In Workshop 1 the client wrote golden questions and named agent candidates. Here you turn them into the things users touch: [assistants](/implementation/key-concepts#assistant) in chat, [document templates](/implementation/key-concepts#document-template) for branded output, and [agent flows](/implementation/key-concepts#agent-flow) for repeatable multi-step work.
It matters for accuracy because the graph only answers well when the assistant speaks the client's language. If people say "engagement lead" and the graph stores a `WORKS_ON_PROJECT` relationship with `role = "Engagement Lead"`, someone has to write that down. Experio cannot guess it.
## At a glance
| | |
| - | - |
| **Purpose** | Configure assistants, AI Instructions, document templates and agent flows that answer the golden questions and deliver the agent candidates |
| **When** | Phase 5 Consume (week 7–8), after the golden questions pass on the validated graph |
| **Duration** | Two sessions of 2 hours: (1) assistants, instructions and templates; (2) agent flow design and review |
| **Attendees** | Experio: FFE (leads), Experio SMEs as needed. Client: PM, super user, one SME per use case, 2–3 pilot users, brand or document owner for templates |
| **Inputs** | Golden question list (`questions.md`) with scores; agent candidate cards from Workshop 1; the client's branded Word/PowerPoint/Excel files; the list of tools pilot users already work in |
| **Outputs** | Assistant specification; AI Instructions text; document templates uploaded; one agent flow sketch per agent candidate, then a published flow; pilot user acceptance plan |
| **Admin pages** | **AI & Agents >** Agent Configuration, AI Instructions, Model Configurations, Agent Flows, Agent Inbox, Copilot, Personas, MCP Servers; **Model & Define >** Document Templates, Template Categories |
Assistants and AI Instructions need **AI Managers** write access. Agent flows and document templates need a **staff** account. See [Roles & permissions](/implementation/roles-and-permissions).
## Before the workshop
**FFE and Experio SMEs:**
* Sort the golden questions into three groups: answered in chat, answered with a document, needs a multi-step workflow. Most questions stay in chat.
* Collect the vocabulary gaps from accuracy validation. Every golden question that failed with root cause "retrieval" usually points at a missing term mapping.
* Read [Assistants](/admin-guide/assistants), [Document Templates](/admin-guide/document-templates) and the [Agent Flows Cookbook](/admin-guide/agent-flows-cookbook).
* Check that the models you plan to use appear in **AI & Agents > Model Configurations**.
**Client homework:**
* The document owner brings the branded Word template (and PowerPoint or Excel, if used) with standard style names: Title, Heading 1–4, Normal, List Bullet.
* Each SME brings one good past example of the output they want (a past-performance write-up, a staffing shortlist).
* Pilot users list the tools they work in all day (Claude Desktop, Cursor, Slack, Google Workspace).
## Agenda
**Session 1**
| Time | Topic |
| - | - |
| 0:00–0:15 | Review golden question scores and the three groups |
| 0:15–0:50 | Assistant design: audience, scope, vocabulary, tools |
| 0:50–1:15 | AI Instructions: firm-wide rules versus assistant rules |
| 1:15–1:50 | Document templates: which outputs, which brand files, retrieval and output instructions |
| 1:50–2:00 | Recap and owners |
**Session 2**
| Time | Topic |
| - | - |
| 0:00–0:45 | Design each agent flow on paper |
| 0:45–1:15 | Build or draft with the Admin Copilot, live |
| 1:15–1:45 | Run with a real input; review the output with the SME |
| 1:45–2:00 | MCP and Integrations for users in other tools; acceptance plan |
## Running the workshop
### 1. Design the assistant
Start with one general assistant. Add a second only when a group of users needs different tools, instructions or models. Ask the room:
* "Who opens this assistant, and what do they ask first?" This becomes the **Subtitle** and **Welcome Message**.
* "What words do you use that the data does not?" List each term and what it means in the graph. These become **Cypher Instructions** and **Resolve Entities Instructions**.
* "Does it need anything beyond the graph, such as documents or external tools?" This decides the **Tool Groups**.
* "Must a user give context before an answer makes sense?" If so, use the **Intent Resolver Prompt Extension** so the assistant asks before it retrieves. Keep the list short.
Model choice per step lives on the **Model Configuration** tab. A common pattern is a smaller, fast model for the router, template matcher, Cypher generation and entity resolution, and a stronger model for the report writer. The router and template matcher models must support function calling.
### 2. Write the AI Instructions
AI Instructions apply in layers: **System > Global > Client > Assistant**. Ask: "Is this true for every assistant, or only this one?"
* **Global**: firm name, tone, citation rules, what not to answer.
* **Client**: rules for one client organization, where the deployment serves several.
* **Assistant**: rules that only make sense for one assistant's audience.
Keep each instruction to one directive. It is easier to switch off one instruction than to edit a long one.
### 3. Choose document templates
Ask: "Which outputs do you copy into Word or PowerPoint today?" For each one, agree the file type, the category, what data to retrieve and how to lay it out. DOCX templates carry Word styles and headers; PPTX templates need named slide layouts including `blank`; XLSX templates are filled in place by an agent flow.
### 4. Design each agent flow on paper first
Take one agent candidate card at a time. Before anyone opens the canvas, fill in these six boxes with the SME:
| Box | Question to ask |
| - | - |
| **Trigger** | "Who starts it, from where? Chat, a button, an upload?" |
| **Inputs** | "What must the person give us? Which of those are required?" |
| **Graph retrieval** | "Which golden question does the first step answer?" |
| **LLM steps** | "What does the model do with the data: rank, classify, extract, write?" |
| **Human review** | "Who checks it before it goes out? What do they change?" |
| **Output** | "What comes out: a chat answer, a DOCX, a filled Excel form?" |
A flow is worth building when the work has fixed steps, a review, or a file output. If a single chat question plus a document template does the job, skip the flow.
Agent flows run from the canvas, the API or chat. They do not run on a schedule. Scheduled data jobs are [Flows](/implementation/key-concepts#flow) under **Process > Flows**, which is a different feature.
### 5. Personas (optional)
[Personas](/implementation/key-concepts#persona) describe who is asking (for example "Partner" or "Analyst") and can gate a conversation until the user supplies context. They are off by default. Use them only when the same assistant must answer differently for clearly different audiences.
### 6. Users in other tools
Ask: "Where do people want to ask questions without opening Experio?" Each user can create a personal API key for the **MCP server** and use Experio from Claude Desktop, Cursor or VS Code. Users connect Google Workspace and Slack for themselves under **Integrations**.
## Worked example: Northbridge Consulting
**Assistant: "Northbridge Knowledge"**
| Field | Value |
| - | - |
| Title / Subtitle | Northbridge Knowledge / "Projects, people, clients and contracts across all three practices" |
| Welcome Message | "Ask about past projects, who has worked where, or what we committed to in a contract." |
| Agent mode | Deep Agent |
| Tool groups | Knowledge Base, Graph Query, Document Processing |
| Cypher Instructions | "Engagement lead means an Employee with WORKS\_ON\_PROJECT role 'Engagement Lead'. A practice is a Practice node; the three practices are Strategy, Technology and Operations. 'Our work for X' means Projects FOR\_CLIENT X. Contract expiry is Contract.expiration\_date." |
| Resolve Entities Instructions | "NBC and Northbridge refer to the firm itself, never to a client. LSH is Lakeshore Health." |
| Models | Small model for router, template matcher, Cypher and entity resolution; large model for the report writer |
**AI Instructions**
* Global: "Refer to the firm as Northbridge Consulting. Cite a project code whenever you name a project."
* Assistant (Northbridge Knowledge): "When listing people, include their practice and title. Dates in the format 12 March 2025."
**Document template: "Past Performance (DOCX)"**, category **Case Study**, answering Q2.
* Retrieval instructions: "Find all projects for the named client: project name, project code, dates, service line, description, engagement lead, and the governing SOW."
* Output instructions: "One section per project: Client Need, Our Approach, Results, Team. Past tense. No more than 250 words per project."
**Agent flow: "Staffing shortlist"** (UC2, owner Sam Whitfield)
Paper design: triggered from chat or the **Launch a flow** menu; inputs role, skills and client; the graph step answers a Q4/Q5-style question; an LLM ranks candidates with reasons; Sam reviews and edits; the output is a DOCX.
```mermaid theme={null}
flowchart LR
S["Start
params: role, skills, client"] --> KG["Knowledge Graph data
employees with matching skills,
past work for the client"]
KG --> R["LLM generate
rank top 5 with reasons"]
R --> H["Human review
Sam approves or edits"]
H --> D["Generate DOCX
bound to review content"]
D --> E["End"]
```
The flow's description, which the chat agent reads to decide when to offer it: "Builds a ranked shortlist of Northbridge consultants for a role, given required skills and a client, for the Resource Manager to review." The **Sub-agent** toggle is on.
## Accelerate with AI
The **Admin Copilot** (**⌘J** or **Ctrl+J**) is docked on the Agent Flows list and editor. Paste the paper design as plain text, for example: "Build a flow that takes a role, skills and a client, queries the graph for matching employees, ranks the top five, has someone review it, and produces a Word document." The copilot drafts the flow as a proposal card, checked on the server. It **proposes only**: you apply the card, save, validate and publish. It can also explain a failed run and propose a targeted fix. See [Admin Copilot](/admin-guide/admin-copilot).
## Entering it in Experio
**AI & Agents > Agent Configuration > Add Assistant.** Fill the Basic, Configuration, Model Configuration, Custom Tool Configuration and Agent Configuration tabs. Details in [Assistants](/admin-guide/assistants).
**AI & Agents > AI Instructions > Create New.** Set scope and order. See [AI Instructions](/admin-guide/ai-instructions).
Check or add the category in **Model & Define > Template Categories**, then **Model & Define > Document Templates > Add Template**: name, type, category, retrieval and output instructions, file. Users pick it from the wand button in chat. See [Document Templates](/admin-guide/document-templates).
**AI & Agents > Agent Flows > New flow**, or start from a template (RFP Response Pipeline, Knowledge Graph Report, RFQ Generation). Drag blocks, connect them, bind inputs, then **Validate** and **Publish**. Bind the DOCX step to the Human review `content` output, not to the draft, so a reviewer's edits reach the file. Turn on **Sub-agent** and write the description. See [Agent Flows](/admin-guide/agent-flows) and the [Flow Blocks Reference](/admin-guide/agent-flows-blocks).
Reviewers act in the chat's flow-run panel, or staff use **AI & Agents > Agent Inbox**.
Personas: turn on `PERSONAS_ENABLED` in **Administer > System Settings**, then create them in **AI & Agents > Personas** ([Personas](/admin-guide/personas)). MCP: each user opens their avatar menu > **MCP Server** and generates a key ([MCP Server](/user-guide/mcp-server)). For flows that call external tools, register servers in **AI & Agents > MCP Servers**.
## Common pitfalls
| Pitfall | Fix |
| - | - |
| Five assistants on day one | Start with one; split only when tools or audiences really differ |
| Vocabulary buried in the system prompt | Put graph terms in Cypher and Resolve Entities Instructions, where the graph steps read them |
| A router or template matcher model without function calling | Every conversation fails at the first step. Pick a model that supports it |
| Building a flow for a one-question job | Use chat plus a document template |
| DOCX bound to the draft, not the review | The reviewer's edits are silently dropped. Bind to `content` |
| Sub-agent flow with no description | The chat agent cannot recognise when to offer it |
| Word template with renamed styles | Use standard style names (Heading 1, Normal, List Bullet) |
| Expecting a flow to run nightly | Agent flows have no schedule. Use **Process > Flows** for data jobs |
## Exit criteria
* [ ] Every golden question is assigned to chat, a document template, or an agent flow
* [ ] Assistant configured; vocabulary written into Cypher and Resolve Entities Instructions
* [ ] AI Instructions entered at the right scope and reviewed by the super user
* [ ] Each document template generates a correct, branded file from a golden question
* [ ] Each agent flow has a signed-off paper design, is published, has run on a real input, and its reviewer has approved one output
* [ ] Sub-agent flows have descriptions and launch from chat
* [ ] Pilot users know how to use MCP and Integrations, if in scope
* [ ] Golden questions re-scored with the final assistant
## Next
Run user acceptance with the pilot users, re-scoring the golden questions as described in [Accuracy validation](/implementation/accuracy-validation). Then continue to [Go-live](/implementation/go-live). Copy-ready worksheets are on [Templates & Worksheets](/implementation/templates).