> ## Documentation Index
> Fetch the complete documentation index at: https://relevanceai-docs-tsp-1431.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Build with Invent
> Invent is a forward deployed agent that builds, manages, and improves the specialist Agents that run your business processes.
## What is Invent
Invent is a forward deployed agent built into Relevance AI. Give it a business process and it works with you to understand how that process runs, including the systems involved and the decisions that require judgment.
It then creates and coordinates the Agents, Tools, Triggers, and Workforces needed to carry out the process, with performance requirements in place before real work begins.
Its job continues after launch. Invent reviews live performance, detects quality problems, and proposes and tests improvements. You bring the process knowledge; Invent builds and manages the Agent team underneath it.
## What can Invent do?
Invent works through your process with you, including its inputs, systems, judgment calls, exceptions, and definition of success. It can read the process back to you, identify missing information, and help determine which parts should become Tools, Agents, or Workforces.
Invent creates the Agents needed for the process, equips them with the required Tools and context, connects the systems where the work happens, and wires the Triggers and Workforces that coordinate the process from end to end.
Invent defines the checks each Agent must pass before launch and samples live runs after deployment. When quality drops, it diagnoses the problem and proposes changes to prompts, Tools, examples, models, or orchestration. It can test those changes against the same requirements before you approve them.
Invent reviews Agent activity, errors, Evals, and performance dashboards to diagnose problems and recommend improvements. Describe a recurring process, and it can help identify where Agents, Tools, or Workforces could automate it.
## Where to find Invent
* **Left sidebar:** Click Invent under Build.
* **Home page:** Click Build with Invent.
* **Floating button:** Click the floating button to open the Invent panel from any page in your project.
## Access controls
### Availability
Invent is available to organization admins by default. An organization admin can turn on Invent for everyone in an individual project from that project's settings.
### Approval controls
Use the approval control beside the message composer to set how Invent handles tool approvals for the current session:
* **Ask first** pauses before tools set to **Ask** or **Always ask** run. Tools set to **Auto** still run without approval.
* **Auto-accept** lets tools set to **Ask** run without approval. Tools set to **Always ask** still require your approval.
Tool permissions save how Invent should handle each action for you in the current project. Set an entire category at once or choose a different setting for an individual tool.
| Setting | What it means |
| -------------- | --------------------------------------------------------------- |
| **Auto** | Runs without asking. |
| **Ask** | Requests approval unless **Auto-accept** is on for the session. |
| **Always ask** | Requests approval every time, even when **Auto-accept** is on. |
Run actions default to **Auto**, create and edit actions to **Ask**, and destructive actions such as deleting or publishing to **Always ask**. Setting a destructive action to **Auto** allows Invent to perform it without another tool-approval prompt.
In the Invent panel, click the three-dot menu, then select **Invent settings**.
Select **Permissions**, configure categories or individual tools, then click **Save**.
These settings control approval prompts, not access. Invent still uses your existing project, asset, and integration permissions. Each user manages their own preferences in each project; organization and project admins cannot use this screen to grant access or set approval rules for other users.
### Enterprise security controls
Invent's access to Relevance AI is scoped to the active project at the API-key and fine-grained authorization layers, not only in the interface. Each internal platform request uses the active project ID and the caller's API key, then passes through the standard authorization checks. Enabling Invent in a sandbox project does not grant it access to assets in a production project.
Within the active project, Invent can use connected integrations such as Salesforce or Gmail with credentials available to that project. The project is the access boundary, so connect only sandbox credentials if the project must remain isolated from production data.
Conversation content is sent to the selected LLM provider to generate responses. Web search is off by default and used only when enabled for a message. Voice is feature-gated and off by default. Memory is off unless the user enables it. Audit data and operational metrics stay within Relevance AI unless the organization has configured enterprise event streaming to a customer-controlled destination.
## How to use Invent
Invent is expensive to run and consumes significant credits. Use it deliberately and keep an eye on your usage.
## Invent features
Describe a process you know end to end, including the outcome, systems, judgment calls, and measure of success. Invent works through it with you before building the Agent team.
You can also choose a task suggested from your project setup, such as creating evals if none exist.
About You tells Invent who you are so it tailors every response to your role and work from the first message — no re-explaining your context each session. It shapes how Invent talks to you and what it assumes about your work, but it doesn't change Invent's role or hardcode any Tool configuration.
It adjusts the level of technical detail, makes domain-aware assumptions, and lets Invent address you by name at natural moments.
To set it up:
In the Invent panel, click the three-dot menu, then select **Invent settings**.
Select **General**, enter your details under **About You**, then click **Save**.
#### Model selection
The model dropdown sets which large language model (LLM) powers your Invent session.
The dollar signs show each model's relative credit cost. Labels such as **Fast**, **Balanced**, **Powerful**, and **Thorough** summarize the trade-off between speed, cost, and capability. The icons beside each model show whether it can work with images and file attachments.
Your plan determines the default model. You can switch models at any time from the dropdown in the Invent panel.
#### Credit usage
The counter beside the model selector shows the credits used in your current session and updates with each interaction.
Past sessions in your conversation history show each task with a status: approved (the action was applied), idle (waiting for your input), or completed. You can review what Invent did and track credit spend across sessions from the history view.
When Memory is enabled, Invent can save and recall details you want it to remember across conversations, such as naming conventions, preferred models, workflows, and recurring constraints.
To view, search, edit, or delete your memories:
In the Invent panel, click the three-dot menu, then select **Invent settings**.
Select **Memory**, manage your memory settings, then click **Save**.
Deleting a memory removes it permanently. Invent will re-learn that information if it comes up again in future conversations.
When Invent hits a platform issue it can't fix, it logs a bug to the Relevance AI engineering team on your behalf. These reports go directly to engineering rather than support, so you won't receive progress updates.
Before setting up a scheduled check-in, add an email or Slack notification channel in [Alerts](/build/invent/invent#alerts) so Invent can notify you about the result.
#### Schedule a check-in
Tell Invent what to review and when. It can suggest an interval based on the Agent's activity.
#### What happens when a check-in runs
Invent reviews evals, errors, and recent activity. If it finds an issue, it sends a summary and suggested fix. If everything looks healthy, it stays quiet.
Add at least one email address or Slack channel so Invent can notify you when a scheduled check-in finds something.
In the Invent panel, click the three-dot menu, then select **Invent settings**.
Select **Alerts**, add an email address or Slack channel, then click **Save**.
Ask Invent to roll an asset back to an earlier version — for example, "Restore the previous version of this tool". Invent copies that version into the current draft, leaving the live published version unchanged until you publish. This restores an earlier version of an existing asset; it doesn't recover one that's been deleted.
Invent always asks you to confirm before restoring, and appends `(restored)` to the version name so the restored state is easy to identify later. For a full guide to version history — including how to restore and rename versions from the builder UI, and how retention works — see [Version history](/build/agents/build-your-agent/version-history).
## Manage Evals with Invent
Invent can build an initial Evals suite, review results, investigate failed Checks, and diagnose performance alarms for Agents and Workforces.
Learn how to create scenarios and reusable Checks, run evaluations, and monitor live performance.
Invent can turn an Agent or Workforce's real usage into a starting Evals suite, including a test set, scenarios, and reusable Checks.
On the **Run** page, select **Create with Invent** from the testing banner. If a recent Agent task is available, you can instead select **Automate testing** beside it so Invent uses that run as the example.
Add any outcomes, edge cases, or requirements that the Evals should cover. A representative task gives Invent useful context about the inputs, expected behavior, and tools involved.
Invent creates the Evals setup and runs the test. Open **Evaluate → Runs** to review the score and see which Checks passed or failed.
Start with a task that represents normal production use, then ask Invent to add edge cases once the baseline Evals are working.
Use **Review with Invent** for an on-demand analysis of a completed Evals run or Performance dashboard. Selecting it starts a fresh conversation with the relevant page as context, including the dashboard name when applicable. Invent summarizes the results and recommends improvements, and you can ask it to compare Checks, explain score changes, identify recurring weak points, or prioritize the highest-impact recommendations.
A review is requested manually; it does not watch performance continuously. Configure an alarm if you want to be notified when a score crosses a threshold.
Target a specific failed or errored Check when you already know which result needs attention. Invent receives the Check name and run page, investigates why it did not pass, and works with you on a fix.
Go to **Evaluate → Runs**, open a completed Evals run, then expand a failed test to see its Check results and failure reasons.
The **Fix** button appears beside each Check that failed or returned an error. Selecting it starts a new Invent conversation focused on that Check.
Invent identifies the likely cause and proposes changes to the Agent, Workforce, or Evals setup. Review the proposed actions before applying them, then rerun the Evals to confirm the Check passes without breaking other behavior.
Use **Review with Invent** for a summary of the entire run; use **Fix** when you want Invent to investigate one failed Check.
#### Turn on auto-diagnosis
Turn on **Auto-diagnosis by Invent** to investigate the first breach and send a diagnosis and proposed fix to the alarm creator's [Invent alert channels](/build/invent/invent#alerts). Invent still shares its findings if the cause is unclear and never applies changes without your approval.
Alarms are configured in an Agent or Workforce's **Evaluate → Monitor** area, not in Invent settings.
Alarms evaluate the scores collected by a Performance dashboard and notify you when an overall score or individual Check crosses the condition you set.
## MCP vs Invent
Invent and the Relevance AI MCP are often compared, but they solve different problems. Invent builds things *inside* Relevance AI from natural language. The MCP connects Relevance AI *to the outside world* — letting external clients call your Tools, and letting your Agents reach into external services.
A forward deployed agent inside Relevance AI. Describe the process you want to run and it builds or edits Agents, Tools, and Workforces for you, landing changes in draft for review.
Reach for it when you want to build, edit, or debug something inside Relevance AI by describing the outcome.
A bridge to other tools. Call your existing Relevance AI tools from external clients like Cursor or Claude with the [MCP Server](/integrations/mcp/mcp-server), or let your Agents call services like Notion and Linear with the [MCP Client](/integrations/mcp/mcp-client).
Reach for it when the Tool already exists and you want to run it from Cursor or Claude, or let an Agent call an outside service.
The two are complementary, not competing. Invent uses the same MCP tool surface under the hood — so a common pattern is to build your Agents and Tools with Invent, then call them from Cursor or Claude through the MCP in your day-to-day work.
## Frequently asked questions (FAQs)
Choose a process you know well and can explain from beginning to end. Tell Invent the outcome you need, the systems involved, the decisions that require judgment, and how you measure success. Invent will work through the process with you before it builds the Agent team.
Invent is available to organization admins by default. An organization admin can turn on Invent for everyone in an individual project from that project's settings.
Invent's access to Relevance AI is scoped to the active project at the API-key and fine-grained authorization layers, not only in the interface. Each internal platform request uses the active project ID and the caller's API key, then passes through the standard authorization checks. Enabling Invent in a sandbox project does not grant it access to assets in a production project.
Within the active project, Invent can use connected integrations such as Salesforce or Gmail with credentials available to that project. The project is the access boundary, so connect only sandbox credentials if the project must remain isolated from production data.
Conversation content is sent to the selected LLM provider to generate responses. Web search is off by default and used only when enabled for a message. Voice is feature-gated and off by default. Memory is off unless the user enables it. Audit data and operational metrics stay within Relevance AI unless the organization has configured enterprise event streaming to a customer-controlled destination.
By default, Invent makes changes in draft so you can review them before they go live. Auto-accept lets Invent apply changes without pausing for each approval, but restoring an earlier version always requires confirmation.
Yes. Invent saves conversations as sessions automatically. Open the **Chats** dropdown in the Invent panel to browse or search previous sessions and continue where you left off.
The Invent panel displays a real-time credit counter that updates with each interaction. The counter tracks the credits used in your current session.