📌 This article is for users of elvex 2.0. Find out which version you're using.
Part 1: Framing Your Problem Statement
Before you start configuring an agent, the most important work happens with the agent builder. A well-framed problem statement to the agent builder is the difference between an agent that works reliably and one that constantly surprises you. The more specific and complete it is, the less you'll need to edit in configurations later.
1.1 Define the Core Job
Start with one sentence: What is this agent's single most important job?
Avoid the temptation to make it do everything. Focused agents outperform general-purpose ones because their instructions, datasources, and tools can all be tuned to a specific outcome.
❌ Too Broad | ✅ Well-Scoped |
"Help the sales team with anything" | "Answer inbound questions about our product pricing and packaging using our pricing sheet and help docs" |
"Summarize stuff for executives" | "Generate a weekly pipeline summary from Salesforce data, formatted as a 5-bullet brief" |
"Be an HR assistant" | "Answer employee questions about PTO, benefits, and onboarding using our HR policy documents" |
Ask yourself:
Who is the primary user of this agent?
What problem does it solve for them, every single time?
What does "done well" look like?
1.2 Identify the Inputs
What information will the agent need to do its job? Think in two categories:
1. What the user will provide at runtime (dynamic inputs)
A customer name or account ID
A question typed into chat
An uploaded file (e.g., a contract to review)
2. What the agent always has access to (static context)
Company context set by your admin
Your personal context (role, preferences, working style)
Datasources you connect to the agent
Being explicit about inputs helps you write tighter instructions and choose the right datasources.
1.3 Map the Data and Tools the Agent Needs to Do Its Job
Every agent needs two things to be effective: information to draw from and actions it can take. Map both before you start configuring.
Knowledge: What does the agent need to know?
Datasources ground the agent in your organization’s actual information. Without them, it relies on general AI knowledge — which can be outdated or fabricated.
Integrations allow the agent to pull in data from third-party tools to help the agent complete its task.
Tips:
Keep it focused. Connecting too many datasources forces the agent to guess which one is the most relevant and where to look first.
Name datasources clearly; the agent reads the name to help understand what’s inside.
For tabular data (spreadsheets, CSVs, databases), the agent can run SQL-style queries which is great for “how many,” “what’s the total,” and “which accounts” questions.
Actions: What does the agent need to do?
Integrations also allow you to push out data created by your agent into third-party tools.
If the agent needs to… | Integration to use |
Send an email | Gmail / Outlook |
Create or update a task or doc | Notion, Linear, Confluence |
Log a call or update a deal | Salesforce, HubSpot |
Post a message | Slack or Teams |
Read or write spreadsheet data | Google Sheets |
Check or create calendar events | Google Calendar |
Tip: Only enable the integrations the agent genuinely needs. Every extra one adds surface area for unexpected behavior.
Example: Post-Meeting Follow-Up Agent
Need | Solution |
Pull relevant account history | Account notes datasource synced from Google Drive |
Log call notes | Salesforce integration |
Send follow-up email | Gmail integration |
Create a follow-up task | Asana integration |
1.4 Define the Desired Output
Be explicit about what a good response looks like. The agent's rules should describe:
Format: Should responses be bullet points? A structured table? A short paragraph? A full document?
Length: Is brevity important, or does the user want thorough answers?
Tone: Professional, casual, concise, detailed?
Citations: Should the agent reference where it found information?
Structure: Are there sections or headers the output should always include?
Examples:
"Always respond with a 3–5 bullet summary. End with a 'Next Steps' section. Use plain, professional language."
"Generate a full written report with an Executive Summary, Findings, and Recommendations section. Cite all datasource documents used."
"Keep responses under 3 sentences unless the user asks for detail. Never use bullet points."
Part 2: Agent Configuration Walkthrough
Once your problem statement is solid, you're ready to configure the agent. Here's every section in the builder, what it does, and a concrete example of when you'd use it.
Summary: Configuration at a Glance
Section | What It Controls | Key Question to Ask |
Name | How the agent is identified | Is the name specific enough to be unambiguous? |
Description | How the agent is discovered | Would a teammate know if this is the right agent for them? |
Context | What the agent always knows | What does it need to know that users shouldn't have to explain? |
Rules | How the agent behaves | What format, tone, and limits should every response follow? |
Model | Intelligence and speed | Does this task need deep reasoning or fast, simple responses? |
Actions | What the agent can do | Does it need to search, analyze data, write code, or create files? |
Datasources | What the agent knows | What internal knowledge grounds this agent in accurate answers? |
Integrations | What the agent can act on | Does it need to take action beyond just responding in chat? |
Visibility | Who can access the agent | Is it ready for my team? The whole org? |
2.1 Name
What it does: Sets the display name of the agent across the platform. This name appears in the agent library, the context strip, and when teammates search for agents to use.
Why it matters: A clear, specific name helps users (and elvex itself) understand what the agent is for at a glance. elvex uses agent names when deciding which agent to suggest for a given task.
Example: "Competitive Battlecard Agent" is immediately clear. "Sales Helper" is not.
2.2 Description
What it does: A short summary of the agent's purpose, visible in the agent library. Also used by elvex when surfacing agent suggestions to users.
Why it matters: The description helps teammates know whether to use this agent for their task — and helps elvex route requests to the right agent automatically.
Example: "Generates real-time competitive battlecards for any prospect's competitor. Uses the Competitor Intelligence datasource and company positioning docs."
2.3 Rules
What it does: Behavioral instructions that shape how the agent responds — its tone, format, length, what it should and shouldn't do. Rules are evaluated on every response.
Why it matters: Rules are the primary way you control the agent's output quality. Without them, the agent will use its best judgment — which may not match your expectations.
Best practices for writing rules:
Use "Always…" / "Never…" / "When X, do Y" phrasing
Be specific about format and length
Call out edge cases the agent should handle gracefully
Example: "Always respond in plain English. Never provide legal advice — if a user asks a legal question, direct them to the Legal team. Always cite the document you pulled information from. Keep responses under 200 words unless the user asks for detail."
2.5 Welcome Message
What it does: Gives users an idea of what their first step should be without having to ask the agent.
Why it matters: If you are building an agent for others to use, the welcome message is a great way to direct others in how to begin their conversation versus the standard How can I help you right now?
Example:
Share the excel you want to analyze
Ask me a question about your company benefits
2.6 Sample Threads
What it does: One click prompts to kick off a conversation with the agent.
Why it matters: If your agent is performing a task that does not need a lot of custom input from the user, Sample Threads are a neat way to shorten time to value.
Example: "Look at the product table and provide charts of the performance of the database." "Make a report covering the last 30 days."
2.7 Actions
What it does: Enables specific capabilities the agent can use during a conversation. Available tools include:
Tool | What It Enables |
Web Browsing | Search the internet and read web pages in real time |
Data Analysis | Query spreadsheets and CSVs with SQL-style logic; visualize results as charts |
(Power Tools) | Run Python, bash, and other code; create files (PDFs, PPTX, XLSX, etc.) |
Drafts | Create and iteratively edit text-based drafts - almost like a Google Doc experience |
Image Generation | The ability to generate an image using the AI provider designated by your administrator. |
Intelligent Search | Enhances your agent’s ability to accurately search and retrieve for an answer across multiple datasources |
Sources & Citations | Informs the agent it should link back to original document when it is retrieving the answer to a question |
Why it matters: Tools dramatically expand what an agent can do — from reading the web to building a full Excel report. Only enable what's needed to keep the agent focused.
2.8 Integrations
What it does: Grants the agent permission to access external tools and services.
Why it matters: Enabling integrations globally does not mean every agent you build will have access to those tools. In order for an agent to have access to an external tool, you must explicitly give it access.
Example: A post-meeting summary agent with Salesforce and Gmail integrations enabled can automatically log call notes to the Salesforce opportunity and send a follow-up email to the prospect — all from one conversation.
2.9 Datasources
What it does: Connects the agent to your organization's specific documents, spreadsheets, websites, or databases. The agent will search and retrieve from these when forming responses.
Why it matters: Datasources are what make an agent accurate. Without them, responses are based on general AI knowledge.
Example: An onboarding agent connected to an HR policy datasource (synced from Google Drive) and a benefits guide datasource can answer "How many PTO days do I get?" with a direct, cited answer — rather than a generic guess.
2.10 Slack or Teams
What it does: Allows users to talk to your agent in a third-party chat application.
Why it matters: You can designate who can work with the agent in Slack or Teams, and what channels the agent can operate within.
2.11 People
What it does: Controls who can see and use the agent. Options typically include:
Private — only you
Shared with specific people or groups
Published to the whole organization
Why it matters: You can control who can see and edit your agent in elvex.
2.12 Advanced - Context Box
What it does: This is where you put domain-specific terminology, product knowledge, company facts, or anything the agent needs to know to do its job well.
Why it matters: Context reduces reliance on the user to explain things the agent should already know. It makes responses more accurate and on-point from the first message.
Example: In an agent from an automotive startup “Instead of Chief Product Officer, when a user types CPO take it to mean certified pre owned unless otherwise specified.”
2.13 Advanced - Enable User Context
What it does: This tells the agent to explicitly inject company and personal context into its work.
Why it matters: Context reduces reliance on the user to explain things the agent should already know. You may or may not want your agent to inject your personal context into its output.
2.14 Advanced - Model & Provider
What it does: Sets which AI model and provider (e.g., OpenAI, Anthropic, Azure) powers this agent. Different models have different strengths, speeds, and costs.
Why it matters: The right model depends on your use case. Lighter models are faster and cheaper; frontier models handle complex reasoning and long documents better.