Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Create an agent

Every Agent Builder task, from your first prompt to configuring, testing, securing, shipping, monitoring, and managing your agents.

This guide is organized by task. Find what you want to do in the list below, then follow the steps.

Task
Go to

Create an agent by describing it in a chat prompt

Start from a ready-made template

Attach files to my first prompt

Bring in an agent I already have

Build agents from my own IDE

Change the agent by chatting

Attach a file to the builder chat

See how full the chat context is

Keep building in a chat from an older version

Start a new chat or open an earlier one

Rename the agent

Ask for help with the agent

See how the agent is wired

See the steps the agent will follow

Open the agent's settings

Change the name, description, output format, or starters

Write or improve the agent's instructions

Set how independently the agent works

Set how deeply the agent thinks

Trade response speed against quality

Choose the models the agent uses

Keep long conversations within the context window

Detect malicious prompts

Give the agent an action it can take

Turn built-in skills on or off

Call an API or run code

Connect an MCP server

Connect an MCP server with client credentials

Connect an MCP server that signs people in with OAuth

Fill the MCP connection form from a config file

Choose whether an MCP tool asks before it runs

Change which MCP tools the agent can use

Edit an MCP connection

Disconnect or remove an MCP server

Finish an MCP connection after import

Connect a SaaS product

Add my own procedure as a skill

Turn a multi-step process into a skill

Hand part of the job to another agent

Give the agent documents to read

Build or attach knowledge from the chat

Attach existing knowledge or a govern graph

Remove knowledge from the agent

Make the agent run on an event

Change or remove a trigger

Store API keys and credentials

Store reusable values

Review the agent's execution plans

See how many credits the agent has used

Try the agent

Build a test dataset

Run an evaluation and read the scores

Apply the suggested fixes from an eval

Create a widget key or API key

Rename or deactivate a key

Limit where the chat widget can run

Verify who is calling the agent

Mask personal data before it reaches the model

Ask for approval before a skill runs

See which models the organization allows

Save a version of the agent

Publish a version so it goes live

Download a saved version

Choose which executor version runs the agent

Run the agent in my own app

Export the agent as a package

Inspect the agent's runs

Score live conversations

Turn on persona or FAQ memory

Find one of my agents

Delete an agent

You need a Qore account with access to Agent Builder, and a project to build in.


Start a new agent

You start every new agent from Create New Agent in Agents: Describe the task. I'll build the agent to do it.

The Create New Agent screen: a prompt box reading Describe the agent you want to build, with a + button and a send arrow, template chips below it, and the Import an existing agent and Build from your IDE links
Create New Agent.

Create an agent from a chat prompt

1

Describe the agent

In Describe the agent you want to build…, write what the agent should do. For example: "Create an agent that answers questions about loan eligibility using our policy documents."

2

Submit the prompt

Select the send arrow.

3

Review the drafted plan

The platform creates and names the agent, such as Loan Eligibility Advisor. It drafts a plan with What it does, Who it serves, Workflow, and the runtime flow, and pre-selects tools. Refine the plan in the Type here… box.

4

Build the agent

When the plan looks right, select Build Agent. The platform builds the agent and opens the builder: the chat on the left, the canvas on the right.

When the build finishes, the chat summarizes what it set up, and says you can test the agent in the Agent Playground or evaluate it under Regression Evals. Each reply shows the credits it used.

Start from a template

Template chips sit below the prompt box, such as Collections Agent. Use the ‹ and › arrows to scroll through them, then select one.

Attach files to the prompt

Select + in the prompt box to attach files to your description.

Import an agent I already have

1

Open Import an existing agent

Below the template chips, select Import an existing agent. The dialog brings in an agent package exported from Platform.

2

Choose the package

Select Click to upload or drag and drop and choose a ZIP file up to 50 MB.

3

Continue

Select Next.

The Import an existing agent dialog: Bring in an agent package exported from Platform, a ZIP upload area (up to 50 MB), and Cancel and Next buttons
Import an existing agent.

An import restores the agent's knowledge into the project and replaces any knowledge the agent already had there. Govern graphs aren't imported. See Export Agent › Knowledge in the package.

Build agents from my own IDE

Below the template chips, select Build from your IDE. Connect to coding agents (MCP) connects your agents to an MCP-capable coding agent, such as Claude Code or Cursor, with an API key. It lists three MCP servers, each with an example command for Claude Code:

Server
What it does

Agent Builder MCP

Create, configure, and manage agents.

Knowledge MCP

Build datasets, ingest documents, and query the knowledge attached to your agents.

Agent Executor MCP

Interact with your deployed agents by invoking them through the trigger API. Choose the Executor version, such as 1.4.

1

Create an API key

Under API keys, select + Add Key. A key is valid only for the current workspace and account, and you can revoke it at any time. A key created here can do only what your role currently allows. Until you create one, the list reads No keys yet.

2

Add a server to your coding agent

Copy the server's example command, and replace <YOUR_API_KEY> with your key. Every request needs the x-platform-key header.

The Connect to coding agents (MCP) dialog: example claude mcp add commands for Agent Builder MCP, Knowledge MCP and Agent Executor MCP with the server addresses blurred, an Executor version menu set to 1.4, the required x-platform-key header, and an API keys section with Add Key and No keys yet
Connect to coding agents (MCP).

Work in the builder

Each agent opens in the builder. The chat is on the left, and the canvas is on the right.

The chat header holds the agent name and four icons: chat history, new chat, Need Help? Report the Issue (?), and Toggle Panel.

The canvas header holds everything else:

The canvas header: the Overview and Flow tabs, a credits meter showing 1,683 credits and $0.84, the Draft Version selector, the Integrate Agent and Export Agent icons, Agent Config, and Try Agent
The canvas header.
Header item
What it does

Overview, Flow

Switch between the agent's two views.

Credits meter

Shows the credits the agent has used, and their cost.

Version selector

Shows the version you're viewing, such as Draft Version.

</>

Integrate Agent: run the agent in your own app.

Export icon

Export Agent: export the agent as a package.

Agent Config

Opens the agent's settings.

Try Agent

Opens the Agent Playground in a new tab.

On a saved version with no unsaved changes, a Download Agent Version icon also appears.

Change the agent by chatting

Type the change in Type here… in the builder chat, and send it. Each reply shows the credits it used, such as 70.3 credits · $0.0351.

Attach a file to the builder chat

Select + beside Type here…, then Attach Files (Upload files from your device).

The + menu in the builder chat, with Attach Files: Upload files from your device
Attach Files.

See how full the chat context is

The ring beside the send arrow shows how much of the chat's context is used. Hover over it for the breakdown: how full it is, such as 11% Full, the tokens used out of the total, such as ~21.1K / ~200.0K Tokens, and the tokens taken by User messages, Assistant messages, Tool calls, and Tool results.

The context usage breakdown: 11% Full, about 21.1K of 200.0K tokens, with User messages 66, Assistant messages 918, Tool calls about 10.0K and Tool results about 10.1K
Context usage.

Continue a chat made in an older version

A chat made in an older version shows This chat was made in an older version. Choose one:

  • Migrate chat — keep building in this chat with its history. Migrating can take a while.

  • Start new conversation — start a fresh chat.

The notice This chat was made in an older version, with Migrate chat and Start new conversation buttons, above the message box
A chat from an older version.

Start a new chat or open an earlier one

In the chat header, select the new chat icon to start a fresh conversation with the builder, or the chat history icon to open a past one. To hide or show the chat, select Toggle Panel.

Rename the agent

Select the agent name at the top of the chat, then Rename Agent. The same menu has Back. You can also rename it in Agent Config › Basic Details.

The agent name menu open, with Back and Rename Agent
The agent name menu.

Ask for help with the agent

Select ? in the chat header. Need Help? Raise a Request opens with the agent filled in.

  • Title — required, up to 200 characters.

  • What do you need help with? — required, up to 2,000 characters.

  • Attachment — optional. Upload up to five PNG or JPEG screenshots, each up to 10 MB.

Select Submit Request. Qore confirms with Issue reported successfully. Thank you for your feedback!

The Need Help? Raise a Request form, with the Agent field filled in, Title and What do you need help with? fields, an Upload button for attachments, and Cancel and Submit Request buttons
Need Help? Raise a Request.

See how the agent is wired

Select Overview. The agent sits in the center, with four cards around it:

  • Triggers — what starts the agent. + Add Triggers.

  • Knowledge — what the agent can read. + Add Knowledge.

  • Skills — the actions the agent can take, each with its type, such as System Skill, and its status, such as Active. View all lists every skill. + Add Skills.

  • Integrations — the connected services. + Add Integrations.

An empty card reads, for example, No knowledge added yet. Each card's count shows how many items it holds.

The Overview tab: the agent in the center, a Triggers card with one Schedule trigger and a Knowledge card with no knowledge on the left, and a Skills card with 31 skills and an Integrations card with no integrations on the right
The Overview tab.

See the steps the agent will follow

Select Flow. It lists the steps the agent may follow at runtime, under What the Agent will do. With no execution plan yet, it shows No Plan Found.

The Flow tab: What the Agent will do, followed by a sequence of steps, each with a short description
The Flow tab.

Shape how the agent behaves

Open Agent Config

Select Agent Config in the canvas header. It opens on Basic Details, with every section in the left navigation. Each section has its own Save, which activates when you change something. Saved changes go to the draft.

Edit the agent details

In Agent Config › Basic Details:

  • Name — required.

  • Description — what the agent does. Required.

  • Output Format — Markdown (the default) or Json. Required.

  • Conversation Starters — the suggested first messages. Select Manage to change them.

Select Save.

Write or improve the instructions

In Agent Config › Instructions, write what the agent does and how. The editor shows your text as formatted markdown. Two controls sit at the top right, and one at the right end of the formatting toolbar:

  • Upload from file — import instructions from a file.

  • Save — save your changes. It stays disabled until you change something.

  • Best practices — open a side panel with the recommended structure: Objective / Goal, Inputs Required, Process / Steps, Outputs / Deliverables, Constraints / Rules, and the optional Tone / Interaction Style and Additional Instructions.

Keep the instructions within 5,000 characters. The counter in the footer shows your length, for example 2,269 / 5,000. Past 5,000 it turns amber and warns that the instructions may overload the context during agent execution, but you can still save. See Instructions.

Select Save.

Set how independently the agent works

In Agent Config › Agent Behaviour, choose Autonomy: Low, Medium (the default), or High. Higher autonomy commits to multi-step plans on its own; lower autonomy confirms more steps first.

Set how deeply the agent thinks

In Agent Behaviour, choose Capability:

  • Auto (the default) — the platform decides based on the input.

  • Intelligent Planner.

  • Next Best Step — hides the Planning Model field.

  • Fast — strips persona, FAQ, CoT validation, and learning cases for lower latency. Only on executor versions that support it.

While Response Mode is Flash, Capability can't be changed.

Trade speed against quality

In Agent Behaviour, choose Response Mode: Advanced (the default, higher quality, slower), Balanced, or Flash. The model fields appear only in Advanced.

Choose the models

With Response Mode on Advanced, under Advanced Controls:

  • Model Family — filter the model lists to one provider family. Agents can use Google and OpenAI.

  • Planning Model — the model that decides what to do, and in what order.

  • Execution Model — the model that carries out the tasks.

A model the AI Gateway can't serve is greyed out. A model the executor version doesn't support is marked Not Supported, and the page can't be saved until you choose another. Select Save.

Keep long conversations within the context window

In Agent Behaviour, turn on Auto-compaction to summarize the conversation as it nears the context window. Set Compaction threshold from 50% to 95%. It appears only on executor versions that support it.

Detect malicious prompts

In Agent Behaviour, turn on Prompt Sanitization. On an executor version that doesn't support it, the toggle is marked Not Supported.

For every behavior setting, see Agent Behavior.


Give the agent skills

Add a skill

1

Open Skills & Capabilities

On the Skills card, select + Add Skills. It opens Agent Config on Skills & Capabilities.

2

Choose a tab

Tab
Holds

All

Every skill from every tab.

Skill Library

Skills by Uptiq that you configure, such as MCP Server, API Call, API Call Advanced, JavaScript Code, and Python Code.

Flows

Flows you can add as a skill.

Agents

Other agents in the project.

MCP Marketplace

MCP servers from the catalog.

Custom Skills

Skills you upload, as a folder with a SKILL.md file.

Pre-attached Skills

The built-in skills, each with Enable.

3

Find the skill

Type in Search Skills…. The search covers the selected tab.

4

Configure and save

Select the skill to open its form. Fill it in, then save. The planner can now choose it when a request needs it.

Turn built-in skills on or off

In Agent Config › Skill & Capabilities, the Pre-attached Skills page has two lists:

  • Active Skills — the skills the agent can use now, each with an Actions menu.

  • Skills you can add — select Enable to make one active.

Every agent has the same 26 built-in skills, such as Show Message, AI Reasoning, Web Search, Knowledge Search, and View Uploaded Files. Skills from other sources appear on the canvas's Skills card instead.

Call an API or run code

In Agent Config → Skills & Capabilities, select Skill Library, then a skill:

Skill
What you configure

API Call

URL, method, authorization, headers, query parameters, and input and output schemas. API Call Advanced adds a request body.

JavaScript Code, Python Code

Name, description, timeout, code, and input and output schemas.

Web Crawler

A Firecrawl MCP connection. Add your Firecrawl API key, or a $secret reference to it, to the pre-filled URL.

Reference stored values as $secret.{name} and $agent.{name} in any field. See Skill Library.

Connect an MCP server

Start from one of:

  • The catalog — open Agent Config › Skills & Capabilities and select MCP Marketplace. Servers are grouped by category, such as Tavily, Firecrawl, and Alpha Vantage API. Select a server.

  • Your own server — select Skill Library › MCP Server.

Both open Create MCP Connection:

1

Enter the connection

Check Connection Name and Remote MCP Server URL, and fill in any field the server needs, such as Api Key. For a key, enter a $secret. reference rather than the key itself: catalog keys are saved in the server URL, which Connectors shows in full. Open Advanced Settings if the server needs OAuth details or headers. Select Next.

2

Authorize, if asked

If Waiting for Authorization appears, sign in to the server in the pop-up and grant the permissions. The dialog continues by itself. If the pop-up was blocked, select Connect to Authorize.

3

Choose the tools

In Tools and Settings, turn off any tool the agent shouldn't use, and set each tool's supervision mode. Then select Create Connection.

The enabled tools become available to the agent. See MCP Tools.

Connect an MCP server that signs people in with OAuth

  1. In Create MCP Connection, expand Advanced Settings.

  2. Copy the URL under Register this Redirect URL, and add it as an allowed redirect URI in the OAuth app on your MCP server.

  3. Enter the OAuth Client ID and OAuth Client Secret from that OAuth app. You can leave them empty if the server supports dynamic client registration; Qore then registers itself.

  4. Select Next, then sign in when Waiting for Authorization appears.

Connect an MCP server with client credentials

For a server that issues a client ID and secret instead of a sign-in page:

1

Open the connection form

Select the server in the catalog, or choose Skill Library › MCP Server. Enter the server's name and URL.

2

Turn on M2M

Expand Advanced Settings and select Authenticate M2M (OAuth client credentials).

3

Enter the credentials

Enter the Token URL (required), the Scope (if your provider needs one), and the OAuth Client ID and OAuth Client Secret (both required).

4

Choose the tools

Select Next. There's no sign-in step: the agent requests its own token each time it connects. Choose the tools, then select Create Connection.

The client secret isn't shown again after saving, so re-enter it when you edit the connection. See MCP Tools.

Fill the MCP connection form from a config file

In Create MCP Connection, select Import Config File and choose a JSON file with an mcpServers object, the format MCP clients use. Qore fills the name, URL, headers, and client ID and secret from the first server in the file. See Import Config File.

Choose whether a tool asks before it runs

Each MCP tool has a supervision mode: Always Ask Permissions, the default, or Allow Unsupervised. Set it in Tools and Settings when you create the connection, or later:

  1. Open the MCP connection, then select Manage Tools.

  2. Change the mode beside the tool. The menu beside Enable All sets every tool at once.

  3. Select Save Changes.

Change which MCP tools the agent can use

  1. Open the MCP connection, then select Manage Tools.

  2. Turn tools on or off with their switches, or with Enable All. Select the pencil beside a description to rewrite it.

  3. To fetch the server's current tool list, select Refresh tools.

  4. Select Save Changes.

Edit an MCP connection

Open the MCP connection and select the pencil. Create MCP Connection opens with the saved name, URL, client ID, headers and M2M settings. Re-enter the client secret if the connection uses one.

Disconnect or remove an MCP server

Open the MCP connection:

  • Disconnect turns the connection off. It asks first, because disconnecting stops all ongoing sessions using the server. Connect turns it back on.

  • Remove permanently removes the MCP server from the agent, after a confirmation.

Finish an MCP connection after import

An agent imported from a package exported without Include MCP configuration has MCP connections without credentials. Each one shows Initialize instead of Connect.

Open the connection, select Initialize, enter the credentials in Create MCP Connection, and continue through the steps.

Add an integration

On the Integrations card, select + Add Integrations, or open Agent Config › Integration Tools & Connectors › Apps & Services. Search with Search integrations…, hover the service's row, and select Connect. Then follow the service's sign-in or credentials form. See Connect an integration.

Add a custom skill

In Agent Config → Skills & Capabilities, select Custom Skills, then upload a folder with a SKILL.md file and up to 10 supporting files. See Custom skills.

Add a flow as a skill

In Agent Config → Skills & Capabilities, select the Flows tab. Each flow has a card with its name and description, a version dropdown, and Add. The dropdown starts on Current Published, which always runs the latest published version of the flow. Choose Version N instead to pin the agent to that version. A pinned flow shows on the agent as Workflow · vN. The draft is never offered. With no flows yet, the tab reads No skills found; select + Create Flow to make one. See Create a flow.

Add another agent as a skill

In Agent Config → Skills & Capabilities, select the Agents tab, then the agent. It attaches straight away, with no form or confirmation, and appears on the Skills card as an Agent skill. It's chosen by its description, so make sure the description says what it's for.


Give the agent knowledge

Add knowledge

1

Open Attach Knowledge

On the Knowledge card, select + Add Knowledge.

2

Choose a category

Choose the category that best matches the content, such as Financial Services › Loan Products.

3

Add the content

Import from Google Drive, One Drive, or SharePoint, or select Click to upload and choose a file up to 10 MB.

4

Done

Select Done. Done stays unavailable until at least one file is added. The knowledge appears on the Knowledge card.

The Attach Knowledge dialog: knowledge categories on the left, with Financial Services and Loan Products selected, and Google Drive, One Drive and SharePoint import options above an upload area with a 10 MB maximum file size
Attach Knowledge.

Turn on the Preprocessing Pipeline for scanned or table-heavy documents. For larger files, use Datasets, which accepts up to 15 MB per file.

Build or attach knowledge from the chat

Ask the builder in the chat. It can:

  • Find what exists — ask in your own words, such as "Do we have anything on our 2026 fee schedule?", then ask it to attach what it found.

  • Create and organize — create a knowledge base or dataset, and rename one or rewrite its description.

  • Add files — add the files you attach to the chat to a dataset, and tell you why a file failed to process.

  • Make files searchable — sync a dataset into a knowledge base. The builder reports the result for each dataset.

  • Connect it to the agent — attach a knowledge base, or detach it.

  • Add flash knowledge — put a few small attached files straight into the agent's context on every run, with no dataset and no wait.

Deleting a file from a dataset, or deleting a knowledge base, affects every agent that uses it. The builder asks you to confirm by name first.

Attach existing knowledge or a govern graph

1

Open Use Existing Knowledge

On the Knowledge card, select + Add Knowledge. Use Existing Knowledge opens on its Knowledge tab.

2

Select it

Find it by name, using Search knowledge if the list is long, and tick it. For a govern graph, pick the entry marked Completed. Knowledge already attached isn't listed.

3

Add it

Select Add. It appears on the Knowledge card. A govern graph shows as a Knowledge Store marked Active.

The Use Existing Knowledge dialog on the Knowledge tab, listing CCU Govern Graph as Completed above failed entries, with a search box, a Create Knowledge button and the Add button
Use Existing Knowledge.

Attaching a govern graph needs the Attach/Detach govern graph on agent capability. Upload the graph in Control Center › Govern › Govern Graph first.

Detach knowledge

  • A govern graph — open its node's ⋯ menu on the Knowledge card, then Unlink.

  • A knowledge base — ask the builder in the chat to detach it.

Detaching removes it from this agent only; the knowledge base and its files stay available to other agents.


Triggers, secrets, and variables

Add a trigger

1

Open Agent Triggers

On the Triggers card, select + Add Triggers, or open Agent Config › Triggers. An agent with no triggers shows No triggers created and Add New Trigger.

2

Fill in the trigger

Choose the Trigger Source, and enter a unique Trigger Name and a Description of when it fires. All three are required.

3

Save

Save the trigger. It appears on the Triggers card with its type, such as Schedule.

The description is read by a language model at runtime to decide whether an event matches. Anchor it to specific values, and keep one condition per trigger.

Change or remove a trigger

In Agent Triggers, select the trigger's row to open its editor. Change the source, name, or description, then select Update Trigger, or Cancel. The row's ⋮ menu holds its other actions.

Store a secret

1

Open Secrets & Variables

Open Agent Config › Secrets & Variables, then select + Add Secret.

2

Fill in the row

  • Name — starts with a letter or underscore, and uses only letters, numbers, and underscores.

  • Type — Text (no spaces) or JSON (must be valid JSON).

  • Value — the secret.

  • Secured — tick it to store the value encrypted. It's off by default.

3

Save

Select Save. Reference it anywhere as $secret.{name}, for example $secret.STRIPE_SECRET_KEY.

The Secrets and Variables page in Agent Config, with the Secrets table and its Text or JSON type menu open, the Variables table, and the Built-in section
Secrets & Variables.

A secret's name can't be changed once saved: delete the row and add it again. To change a secured value, untick Secured (this clears the value), enter the new value, tick Secured again, and save.

Store a variable

In Secrets & Variables, select + Add Variable. Enter a Name, an optional Description, and an optional Default Value, then select Save. Reference it as $agent.{name}, for example $agent.BASE_URL.

The Built-in section lists values the platform provides, such as $auth.token, the verified caller's token. It resolves only in MCP connection configurations.

Review the agent's execution plans

Open Agent Config › Execution Plans, under Agent Setup. Each row is a reusable execution plan, with its Name, When to use, number of Steps, and Actions. Select Add Execution Plan to write one. Execution plans also power the Flow tab. See Execution Plans.

See how many credits the agent has used

Read the credits meter in the canvas header, such as 1,683 credits and $0.84. Hover over it for the breakdown.


Test and improve

Try the agent

Select Try Agent in the canvas header, or the ▷ icon on the agent's row in My Agents. The Agent Playground opens in a new tab, where you can chat with the agent. You need the Trigger Agent capability, and the agent needs a widget key.

Build a test dataset

1

Open Regression Evals

Open Agent Config › Regression Evals. On the Datasets tab, select New dataset.

2

Choose where the test cases come from

  • Write them yourself — start with an empty dataset and add each question and expected answer by hand.

  • Generate with AI — set the Number of test cases (up to 50, default 10), and optionally What should they focus on?, such as "Edge cases around refunds and partial payments". The cases are written from the agent's instructions.

  • Import from a document — a .pdf, .xlsx, .xls, or .md file up to 10 MB that lists test cases or describes scenarios. Up to 500 test cases.

3

Review and generate

Enter a Name (up to 120 characters), choose how you will add test cases, and select Create dataset, Create and generate or Create and import. A dataset made from AI or a document shows Adding test cases until its test cases are ready; you can leave the page meanwhile.

The Regression Evals page in Agent Config on the Datasets tab, listing datasets with their status, test case count and creation time
Regression Evals, on the Datasets tab.

Each agent can have up to 100 datasets. A dataset's row menu has View, Edit details and Delete. Each test case holds a question, an expected answer, and optional documents.

Run an evaluation

1

Start the run

Open a dataset with at least one test case, then select Run Evaluation. Qore confirms Evaluation run started.

2

Follow it

On the Runs tab, each run shows its status, progress, and Passed, Failed, and Avg score. Select Stop to end a running evaluation; completed results are kept.

3

Read the scores

Open the run. Each test case is scored 0–10 by an LLM, with its reasoning. 7 or above passes. Each result shows the question, the expected answer, the agent's actual answer, and the scorer's reasoning.

The Regression Evals page on the Runs tab, with a search box and two completed run cards showing passed and failed counts and an average score
Regression Evals, on the Runs tab.

Apply suggested changes

When a run completes with failures or an average score below 7, Qore suggests changes, each in a category: Instructions, Skills, Knowledge, Guardrails, or Other.

1

Open the suggestions

Select Apply to Agent on the completed run.

2

Choose what to apply

Tick the suggestions you want. Edit a title, category, or description if needed.

3

Apply

Select Apply. Agent Config closes, and the builder agent makes the changes in the chat panel.

Run the evaluation again to check the result. See Regression Evals.


Secure the agent

Create an access key

1

Open Access & Security

Open Agent Config › Access & Security, under Credentials & access, then select + Add Key.

2

Name it and choose the type

Enter a Key Name, such as Production Widget, and choose the Key Type:

  • Widget Key — for the embedded chat and the TypeScript SDK. Safe in browser code; works only from the platform and your allowed domains.

  • API Key — a key and secret pair for server-to-server calls.

3

Generate and copy

Select Generate Key. Copy the key, and for an API key its API Secret, before you close the dialog. They can't be retrieved later.

The Add New Key dialog, with a Key Name field, a Key Type choice of API Key or Widget Key, and the Generate Key button
Add a key.

To copy an existing key later, select the copy icon beside its masked value in the Key column.

Rename or deactivate a key

On the key's row, open ⋯, then Rename or Deactivate. Deactivating can't be undone. There's no rotate action: add a new key, move your integrations to it, then deactivate the old one.

The Access Keys table with a row's menu open, showing Rename and Deactivate
A key's menu.

Limit where the widget can run

In Access & Security, under Allowed Domains, select + Add New Domain, enter the URL or domain in Enter URL, and select Save. Remove an entry with its trash icon.

  • With the list empty, the widget runs nowhere outside the platform.

  • The list applies to widget keys only.

  • Enter each domain exactly as the browser sends it. A mismatch in protocol, subdomain, or port is refused.

The Access and Security page in Agent Config, with the Access Keys table showing one active widget key and the Allowed Domains list below it
Access & Security.

Verify who is calling the agent

1

Register an app with your identity provider

Register an application in Microsoft Entra ID, or create an OAuth client in your Google Cloud project.

2

Turn on Authentication

In Agent Config › Authentication, turn on Enable Authentication, and choose the Authentication Provider.

3

Enter the provider's details

  • Microsoft Entra ID — the Directory (tenant) ID and Application (client) ID, from the app registration's Overview page.

  • Google — the OAuth client ID, ending in .apps.googleusercontent.com.

4

Save

Select Save.

Authentication configured with Microsoft Entra ID, showing the Directory tenant ID and Application client ID fields.
Authentication with Microsoft Entra ID.

Your app then sends the signed-in user's token with each request. Requests without a token still run; requests with an invalid token are rejected. See Integrate an agent › Verify who's calling.

Mask personal data

1

Turn on Sensitive Data Masking

Open Agent Config › Guardrails. Select the Sensitive Data Masking row to open its modal, and use the row's ⋯ menu to Activate it.

2

Choose a provider

Choose AWS Comprehend, Azure Text Analytics, Google Cloud DLP, Regex, or Off. The provider decides which entity types you can pick.

3

Set the entity rule

  • Deny list — mask only the entity types you select.

  • Allow list — pass the selected types through, and mask every other detected type.

Use Select all, Clear, or Clear group, and Search to find a type.

4

Save

Select Save.

The PII masking configuration screen, with a provider dropdown
PII masking.

Masked values are replaced with tokens before the model sees them. Tokenization needs a compatible executor version; otherwise the entity list is read-only. See PII masking.

Require approval before a skill runs

1

Turn on Tool Execution Approval

In Agent Config › Guardrails, select the Tool Execution Approval row to open its modal, and use the row's ⋯ menu to Activate it.

2

Choose the skills

On Edit Tool Permission, turn on approval for each skill that needs it, or turn on Enable for all skills to cover every current and future skill. Search skills filters the list.

3

Save

Select Update Tool Permission, or Cancel.

When the agent reaches one of those skills, it pauses and asks the user. Allow runs the skill; Deny blocks it, and the agent continues without it.

The chat pausing on a Web Search tool call, asking the user to Allow or Deny
A skill waiting for approval.

See which models are allowed

In Agent Config › Guardrails, Allowed LLM Families and Allowed LLM Providers show as Always enforced. Your organization sets them in the AI Gateway; they can't be changed per agent.

The Guardrails list, with Name, Type, and Status columns for each guardrail
Guardrails.

Ship and run

Save or switch versions

Changes save to the draft. To keep the current state as a version you can return to, select Draft Version in the canvas header. See Runtime & Versioning.

Publish an agent

Publishing makes a saved version the agent's live version, the one new conversations run when nobody asks for a specific version. Publishing needs approval from a second person, so a request goes to Publish Review and the version goes live once it is approved.

  1. In the canvas header, select Publish, next to Try Agent.

  2. A window titled Publish Agent opens, and shows when the agent was last published. With Draft Version selected it reads Send your latest changes for approval? They are saved as a new version and go live once approved. With a saved version selected it reads Send version v2 for approval? It is marked as published once approved.

  3. Select Publish.

Two checks can appear first:

  • Unpublished knowledge. If knowledge attached to the agent hasn't been published, a window titled Publish Agent - Knowledge warns Some knowledge linked to this agent hasn't been published yet. People using this agent may get answers drawn from this content. Select Cancel, or Publish Anyway to continue. See Knowledge.

  • A request is already waiting. If another version is waiting for review, the window asks Replace the pending request? and Another version of this agent is waiting for review. Replace it? Select Replace to withdraw the old request and send the new one.

The button then reads Pending Approval. The version that is already live keeps running until the new one is approved. After approval the window reads Agent Published, and the approved version is live. A reviewer's rejection leaves the live version as it was, and you can submit again.

To publish, you need edit access to the agent and a role with Request agent publish. Developer, Account Admin, and Organisation Admin have it. Approving needs Approve agent to publish (with its tier), which Approver, Account Admin, and Organisation Admin have. Nobody can approve a request they raised themselves.

Why you may not see an approval step. Approval is a separate setting for each kind of publish. If the count for agent versions is 0 in your account, Publish takes effect at once and no request is created, even when other kinds, such as knowledge or apps, still need review. An administrator changes the count through the API. See How approval works.

Download a saved version

On a saved version with no unsaved changes, select Download Agent Version (the download icon) in the canvas header. The version's file downloads to your browser.

Choose the executor version

The executor version is the runtime that runs the agent. Change it in either place:

  • Agent Config › Agent Behaviour › Runtime & Versioning under Advanced Controls, then Save.

  • My Agents: select ••• on the agent's row, then Executor Version, and choose a version. The current one is ticked. 1.0 is marked Deprecated.

My Agents with the ••• menu open on an agent: Export Agent, Executor Version and Delete Agent, with the Executor Version submenu listing 1.4 (ticked), 1.3, 1.2, 1.1 and 1.0 Deprecated
The agent's ••• menu, with Executor Version open.

Run the agent in my own app

Select Integrate Agent (</>) in the canvas header. Choose Embedded Chat for a chat widget, or Agent API to call the agent from your code. For every step, see Integrate an agent.

Export the agent

1

Open Export Agent Package

Select the export icon in the canvas header, or ••• › Export Agent on the agent's row in My Agents. Both open the same dialog. You need the Export Agent capability.

2

Choose what to include

  • Include secrets (API keys, tokens, credentials) — off by default.

  • Include MCP configuration — off by default. Off, MCP connections still export, without their credentials; re-enter them with Initialize after import.

  • Include integration skills — off by default.

3

Export

Select Export Package. Qore emails you a download link when the package is ready. The link expires; export again for a new one.

The Export Agent Package dialog: Export this agent's logic, configuration, tools, and contextual data into a single portable package, with Include secrets, Include MCP configuration and Include integration skills checkboxes, and Cancel and Export Package buttons
Export Agent Package.

Monitor in production

Inspect the agent runs

Open Control Center › Observe › Runs & Traces. Each row is one run, from the widget, the API, or a trigger, with its Execution ID and Conversation ID. Select a row to see the skills chosen, their results, guardrail results, and step timing. See Runs & Traces.

Score live conversations

Online Evals is part of Control Center › Evals. Open Agent List, assign one or more evaluators to the agent, and turn on its Online eval toggle. Each scored conversation appears in Online Eval Runs, with the reasoning behind its score. Scoring runs in the background and never affects the live conversation. See Online Evals in Control Center.

Turn on persona or FAQ memory

Open Agent Config › Reinforcement Learning (Preview):

  • User Persona — turn on Enable User Persona to remember what each user has asked about, preferred, or corrected. You can delete a stored memory.

  • FAQ — turn on Enable FAQ to use stored FAQs and their answers when a question matches.

  • Learning Cases — review corrections from production. Each one needs a person's approval before it's used.

Both memories are off on a new agent, and aren't used while Capability is Fast.


Manage your agents

My Agents, below the prompt on Create New Agent, lists your agents. Each row shows the agent's name, description, and when it was last updated, then ▷ (Try Agent) and the ••• menu.

Find one of my agents

  • Type part of the name in Search agent….

  • Use the list and grid icons to switch how the agents are laid out.

Delete an agent

Select ••• on the agent's row in My Agents, then Delete Agent.


Common issues

  • The agent ignores attached knowledge. Check that Knowledge Search is active, that the instructions tell the agent to consult its sources, and that new files show Synced.

  • The agent doesn't use a skill. Make the skill description specific, and refine the instructions to say when to use it.

  • A skill fails at runtime. Check its credentials and endpoint, then inspect the run in Runs & Traces.

  • An MCP server's tools all fail as unauthorised. The server may expect client credentials. Turn on Authenticate M2M and reconnect.

  • An MCP connection fails at Next on an OAuth server. The server may not support dynamic client registration. Enter the OAuth Client ID (and secret) from its OAuth app under Advanced Settings.

  • The MCP sign-in pop-up didn't open. Your browser may have blocked it. Select Connect to Authorize.

  • An MCP server shows Initialize instead of Connect. It was imported without its credentials. Select Initialize and enter them.

  • An MCP server's API key shows in Connectors. Catalog keys are saved in the server URL. Edit the connection and enter a $secret. reference instead.

  • I don't see the canvas. You may still be in the planning view. Select Build Agent.

  • Agent Behaviour won't save. A selected model or mode isn't available. Choose one that isn't greyed out or marked Not Supported.

  • The chat says it was made in an older version. Select Migrate chat to keep its history, or Start new conversation.

  • The Export Agent icon is disabled. You need the Export Agent capability.

  • I lost an API secret. It's shown only once. Create a new API key, move your integration to it, and deactivate the old one.


Last updated

Was this helpful?