The self-improvement feature allows agents to update their own rules during a conversation, making it easier to refine and improve your agent iteratively without leaving the chat interface.
What is the self-improvement feature?
Self-improvement is an optional feature that gives your agent the ability to modify its own rules based on feedback and examples you provide during a conversation. When enabled, the agent gains access to a special tool that allows it to propose changes to its configuration, which you can review and approve.
This feature is designed for agent creators and editors who want to iteratively refine their agent's behavior through natural conversation rather than manually editing the configuration.
How self-improvement works
When self-improvement is enabled:
The agent can propose rule changes: As you chat with your agent and provide feedback or examples, you can ask it to update its rules based on what you've discussed
You maintain control: The agent will always ask for your approval before making any changes to its rules. You'll see the proposed changes and can choose to accept or reject them
Changes apply to future conversations: Once you approve updated rules, they become part of the agent's configuration and will be used in all subsequent conversations
Each conversation is independent: The self-improvement feature does not allow the agent to learn from other conversations or share information between different conversations. It only allows updates within a single conversation that you explicitly approve
When to use self-improvement
Self-improvement is particularly useful when:
Testing and refining a new agent: You can quickly iterate on rules as you discover what works and what doesn't
Adjusting formatting or tone: If you provide examples of how you want the agent to format responses, you can ask it to update its rules to match
Adding new constraints or guidelines: When you realize the agent needs additional behavioral guidelines, you can add them conversationally
Fine-tuning based on real usage: As you use the agent and notice areas for improvement, you can refine the rules on the spot
What self-improvement is not
It's important to understand the limitations of this feature:
Not cross-conversation learning: The agent does not learn from other users' conversations or remember information from previous conversations
Not automatic improvement: The agent will only propose changes when you explicitly ask it to update its rules
Not a replacement for intentional design: Self-improvement works best when you have a clear vision for your agent and use it to refine that vision, not as a substitute for thoughtful initial configuration
How to enable self-improvement
Self-improvement is only available to users with Editor or Creator permissions for an agent.
To enable self-improvement:
Open a conversation with your agent
Look for the Self-Improvement toggle in the left sidebar, above your recent conversations
Toggle it to the on position
The feature is now enabled for this conversation. You can ask your agent to update its rules based on the conversation context.
Using self-improvement in practice
Here are some examples of how you might use self-improvement:
Example 1: Updating formatting
You: "I'd like you to format dates as MM/DD/YYYY instead of the way you're currently doing it. Can you update your rules to reflect this?"
Agent: Proposes rule change for you to review
Example 2: Adding constraints
You: "I noticed you sometimes make assumptions about what I'm asking for. Can you update your rules to always ask clarifying questions when my request is ambiguous?"
Agent: Proposes rule change for you to review
Example 3: Refining tone
You: "The responses you've been giving in this conversation are exactly the right tone - professional but friendly. Can you update your rules based on the examples from this conversation?"
Agent: Proposes rule change for you to review
Best practices
Be specific when requesting changes: The more specific you are about what you want to change and why, the better the proposed rule updates will be
Test the changes: After approving rule updates, test the agent with a few examples to make sure the changes work as expected
Review proposed changes carefully: Always read through the proposed rule changes before approving them to ensure they match your intent
Use alongside manual editing: Self-improvement is a tool to speed up iteration, but you can still manually edit your agent's configuration at any time
Document major changes: Consider keeping notes about significant rule changes so you can track your agent's evolution over time