Introduction
The Learning feature helps your agents improve over time. When an agent completes a task particularly well, you can capture that success as a “learning” that guides future runs.What is Learning?
Learning allows you to tell an agent: “This is how you should handle this type of task.” By marking a successful run as a learning, you help the agent understand what good looks like. Think of it as positive reinforcement—showing the agent an example of excellent work so it can replicate that success.Why Use Learning?
Improve Consistency
When an agent handles a task perfectly, capture that approach so future runs follow the same pattern.Handle Edge Cases
If an agent successfully navigates a tricky situation, save that learning so it knows how to handle similar cases.Refine Over Time
As you work with an agent, you can continuously improve it by adding learnings from its best performances.Reduce Instruction Complexity
Instead of writing detailed instructions for every scenario, let the agent learn from examples of good work.How to Create a Learning
From Past Runs
- Go to your agent’s run history
- Find a run that completed the task excellently
- Click the Create Learning button at the top of the run
- The agent will now reference this run for guidance
From Current Runs
While viewing a run in progress or just completed:- If the run is going well, look for the Create Learning option
- Click to save this run as a learning
- Future runs will benefit from this example
How Learning Works
When you create a learning:- The agent analyzes what made that run successful
- It extracts patterns and approaches used
- Future runs reference this learning for guidance
- The agent applies similar approaches to new tasks
Best Practices
Choose Representative Examples
Select runs that:- Completed the task correctly
- Handled the typical case well
- Demonstrate the approach you want
- Produced the output quality you expect
Avoid Edge Cases Initially
Don’t make unusual situations the learning:- Start with standard, successful runs
- Edge cases can create unexpected patterns
- Build a foundation with typical examples
One Learning at a Time
Each agent carries only one learning:- Creating a new learning replaces the old one
- Choose your best example
- Update when you find a better example
Review Results
After creating a learning:- Monitor subsequent runs
- Verify the learning improves performance
- Adjust if results aren’t as expected
Real-World Examples
Example 1: Email Responses
Your agent drafts customer responses. One response was particularly well-written—professional tone, addressed all concerns, and followed brand guidelines perfectly. Create a learning from this run. Now future responses will follow this high-quality example.Example 2: Data Processing
Your agent extracts data from invoices. One run correctly handled a complex invoice with multiple line items and applied the right categorization. Create a learning from this run. The agent now knows how to handle similar complex invoices.Example 3: Report Generation
Your agent creates weekly reports. One report had exactly the right structure, formatting, and level of detail. Create a learning from this run. Future reports will follow this successful template.Limitations
Single Learning
Agents carry only one learning at a time:- New learnings overwrite previous ones
- Choose your best example carefully
- You can always update the learning later
Not Visible in Instructions
The learning influences behavior but:- Isn’t shown directly in the instructions
- Works behind the scenes
- Complements your written instructions
Task-Specific
Learnings work best when:- The task is similar to the learned example
- Inputs are comparable
- Expected outputs are consistent
Things to Know
- Learning helps agents improve through positive examples
- Only one learning can be active per agent
- Creating a new learning replaces the previous one
- Learnings work alongside your written instructions
- Choose runs that represent your best outcomes