Create a Self-Learning Agent

A self-learning agent improves its future work by saving reusable knowledge as project skills. It does not retrain its underlying AI model or change its own core instructions.

As the agent works, it can:

• Create a skill for a new recurring workflow, domain, or data source.
• Update an existing skill when it discovers a reliable convention, lookup, or correction.
• Use those skills in future conversations, so your team doesn't have to explain the same thing twice.

The public Skill Curator skill provides this behavior. It is an always-on meta-skill that reviews what the agent learned during a conversation and curates the agent's project skill library.

Before you start

You need:

1. An agent in the target project.
2. Permission to manage that project's skills. Creating and editing skills is limited to project administrators.
3. The Skills Manager toolkit attached to the agent.
4. Create Skill and Edit Skill enabled in the Skills Manager toolkit.

The read-only skill discovery tools are enabled automatically. The agent can't self-curate until the two write tools are enabled.

Informação

Installing a library skill copies it into your project. You can customize your copy without changing the public library skill for other users.

Configure the agent

1. Open the agent's Skills tab

   Open the project, select the agent you want to improve over time, and open its Skills tab.

2. Add the Skills Manager toolkit

   If the Skills tab shows that the Skills Manager toolkit isn't installed, install it and save the agent. Skills need this toolkit at runtime.

3. Allow skill curation

   Open the agent's Toolkits tab, select Skills Manager, and enable:

   • Create Skill (skill_create)
   • Edit Skill (skill_edit)

   Save the agent after changing the toolkit settings.

   Informação

   The agent creates and updates project-scoped skills only. It can't edit public library skills directly.

4. Install the Skill Curator skill

   Return to the Skills tab, then:

   1. Select Browse Library.
   2. Search for Skill Curator.
   3. Install it into the project.
   4. Confirm it appears in the agent's assigned skills and is active.
   5. Save the agent if prompted.

   Skill Curator is Always On. It is included on every conversation turn so it can notice useful lessons as they happen.

aviso

Skills are charged at 1 credit per tool call invoked. See the credits documentation for details.

Tailor Skill Curator to your use case

The public Skill Curator is a general starting point. After you install it, edit your project's copy to match your agent's job, data sources, and business process.

For example, an agent that analyzes Smart Tables can be instructed to:

• Curate only facts verified from the selected Smart Tables.
• Create or update skills only for stable table names, field definitions, join keys, approved filters, and report logic.
• Avoid creating a skill for one-off questions or temporary data values.
• Update an existing domain skill before creating a similar one.
• Never save row-level customer data, ticket text, credentials, or other sensitive information in a skill.

Keep the original curation workflow, but replace broad examples and triggers with the conventions that matter for your agent. Review the edited skill after the first few conversations and refine it as the workflow becomes clearer.

How the agent learns

Before sending its final response, the Skill Curator checks the conversation for reusable information. It creates or updates skills in these cases:

Situation	Example of what is captured
The agent retries a tool call with corrected input	"For customer lookups, use account ID rather than email."
The agent probes a source to discover its structure	"Monthly reports are stored under fyYYYY/q<n>-<topic>/."
A user corrects the agent	"Use the production workspace, not the sandbox, for this report."
A new, non-obvious domain fact is established	"A pending invoice isn't collectible until it has a billing contact."
A repeatable workflow has no existing skill	"Create a weekly customer-health report."
A new data source or integration is explored	Connection details, key identifiers, schema highlights, lookup keys, and known gotchas.

Before creating a skill, the agent searches the project's skill library. If a relevant skill exists, it updates that skill instead of creating a duplicate.

Example: learning from Smart Tables

The scenario

A customer-success team has three Smart Tables in its project:

Smart Table	Purpose	Example fields
Customer Accounts	Master account record	account_id, account_name, owner, plan, status
Product Usage	Daily product activity	account_id, activity_date, active_users, feature_events
Support Tickets	Customer support history	account_id, ticket_id, created_at, priority, state

The team wants one agent to answer questions like "Which accounts are at risk this week?" without someone explaining the table relationships, field meanings, and reliable filters every time.

Configure Smart Table access

1. Create the Smart Tables

   Create the tables in your project and add representative, non-sensitive data.

2. Add the Smart Table Manager toolkit

   Open the agent's Toolkits tab and add Smart Table Manager.

3. Limit table access

   Under Table Access, select Access specific tables only, then choose Customer Accounts, Product Usage, and Support Tickets.

4. Enable read tools

   Enable the read tools the agent needs, especially Describe Table and Query Smart Table. For an analysis-only agent, leave write tools (adding, updating, or deleting rows and columns) disabled.

5. Keep skill curation enabled

   Keep Skills Manager with Create Skill and Edit Skill enabled, and Skill Curator installed, then save the agent.

Dica

Selecting specific tables is safer than giving the agent access to every Smart Table in the project. It also keeps the skills the agent creates focused on the intended business domain.

First prompt: let the agent discover the data

Send the agent a prompt such as:

Analyze our customer health using the attached Smart Tables. Identify accounts at risk when they are active customers, had no product usage in the last 30 days, or have an open high-priority support ticket. Explain the criteria and list the affected accounts.

To answer, the agent may inspect the available tables, describe their columns, and query them. It can learn and save facts such as:

• account_id is the shared key across all three tables.
• status = active identifies current customers.
• activity_date is the field used to calculate the 30-day usage window.
• state = open and priority = high identify urgent unresolved tickets.
• A customer-health report joins account information to usage and ticket data by account_id.

The Skill Curator adds these verified conventions to an existing related skill, or creates a focused one such as customer-health-smart-tables.

Follow-up prompt: reuse the learned skill

After reviewing the new skill for accuracy, send a related request:

Create this week's customer-health summary. Show active accounts with no usage in the last 30 days, open high-priority tickets, and the account owner who should follow up.

The agent loads the learned skill and applies the known tables, join key, filters, and report definition. It only needs to query fresh data instead of rediscovering how the Smart Tables are structured.

What to review after the first run

Open the generated or updated skill and confirm it records only stable, reusable information:

• Table names and their purpose.
• The relationship key (account_id).
• Field definitions and supported status values.
• The approved risk criteria.
• The required read tools and query sequence.
• Meaningful limitations, such as incomplete data or delayed updates.

aviso

Don't let a skill record customer rows, support-ticket content, account notes, or other data that is sensitive, temporary, or not needed to run the workflow.

Review and governance

Self-learning is only useful while the recorded knowledge stays accurate. Review the agent's skills regularly, especially during the first few weeks.

• Keep one skill per clear domain or workflow, and merge overlapping skills.
• Remove temporary facts, test IDs, and expired process instructions.
• Verify that tool names, field names, and paths are correct.
• Use short, lowercase, hyphenated skill names, for example customer-health-reporting.
• Give each skill a specific description and relevant trigger phrases.
• Disable a skill immediately if it contains incorrect or obsolete guidance.
• Never place secrets, passwords, API tokens, personal data, or private customer content in a skill.

You can configure the agent to curate silently, but the people responsible for the project should still own the quality and privacy of the resulting skill library.

Related

Agent Skills Overview

Learn how skills are discovered, activated, and combined.

Getting Started with Skills

Enable skills and create your first custom skill.

Skills Library

Browse pre-built skills, including Skill Curator.

Smart Table Manager

Give agents structured access to your Smart Tables.