Capture the interaction
Connect the conversation and agent activity with application context and human input. Understand what happened before deciding what to learn.
Solutions
Self-improving agents
Learn from user corrections, successful workflows, and the work people complete when an agent misses a step.
Any agent frameworkOpen protocolsSelf-hostableBuilt for your product
A WORKFLOW TO START WITH
A correction today can improve the next attempt.
The agent’s first response is only part of the story. What a person corrects, changes, or does next can reveal a better way to complete the task.
The request
“Check the policy exception before asking me to approve.”Illustrative scenario
The work
Intelligence connects the correction to the surrounding interaction. Repeated patterns become evidence for a reusable lesson, with the source conversations available to review.
The result
Turn the lesson into a Skill. Evaluate it against the task and make the improved approach available when the agent needs it again.
Find reusable lessons
Identify corrections and successful approaches. Turn the evidence into Skills and procedures your agents can use again.
Evaluate the next attempt
Define the behavior you want to improve. Use the source interaction to shape evaluations and check whether the lesson helps on the next task.
Improve the agents you already have.
You do not need to replace your model or rebuild the application to learn from its use. Connect real interactions, organize the learning context, and make useful experience available to your agents.
Explore the documentationStart with a recurring correction.
Choose a task, define what a better attempt looks like, and bring the interactions that show where people intervene. Use that evidence to shape the first Skill.
Explore Automatic LearningPut it to work with your agent.
Copy the setup prompt into your coding agent to start building with CopilotKit.
