Skip to main content

​
Enter the Development Environment

After initializing your agent with bunx blink init, start the development environment: 
blink dev
This opens Blink in your terminal where you’ll spend your time building and testing your agent. The development environment is where Edit Mode and Run Mode come together.

​
Edit Mode (Ctrl+T)

Edit Mode is where you define your agent’s behavior. Think of it as the blueprint for how your agent thinks, acts, and responds. It has the autonomy and access to create your agent’s prompt, give it the tooling you define, and debug how the agent works. Edit Mode In Edit Mode, you can:
  • Describe your agent’s personality and communication style
  • Define what tasks it should excel at
  • Set guidelines for how it should approach problems
  • Specify tools and integrations it should use
  • Add examples of ideal behavior
  • Debug and refine how your agent interprets instructions
Edit Mode has full context of your past chats in both Edit Mode and Run Mode. This gives it optimal context for understanding and debugging how your agent behaves.
Example prompts for Edit Mode: 
You are a research agent that excels at gathering information from multiple sources,
fact-checking claims, and delivering structured reports. Always cite your sources
and highlight conflicting information when you find it.

When planning features, always break down the work into phases, identify risks early,
and ask clarifying questions before proposing solutions. Favor iterative approaches
over big-bang releases.

​
Run Mode

Run Mode is where you use your agent for actual work. This is standard chat interaction where you give your agent tasks and it executes based on the behavior you defined in Edit Mode. Run Mode In Run Mode, you can:
  • Give your agent specific tasks
  • Test how it responds to different scenarios
  • Iterate on real work
  • See how well your Edit Mode instructions work in practice
Run Mode only has context of chats sent and received by your agent in Run Mode. It does not have access to your Edit Mode conversations, which could be confusing to the agent if you reference them.
Example prompts for Run Mode: 
Research the latest updates to React Server Components and summarize
the key changes from the previous version.

Plan out how we should implement user authentication in our app.
Break it down into phases and identify potential risks.

​
The Build-Test-Refine Loop

The most effective way to build agents is through rapid iteration:
1

Define behavior in Edit Mode

Start by describing how you want your agent to behave. Be specific about its role, responsibilities, and approach to problems.
2

Test in Run Mode

Switch to Run Mode (Ctrl+T) and give your agent real tasks. See how it performs with your current instructions.
3

Refine based on results

Switch back to Edit Mode and adjust your instructions based on what worked and what didn’t. Add examples, clarify guidelines, or adjust personality traits.
4

Repeat

Keep iterating. The best agents are built through continuous refinement based on real usage.

​
Tips for Effective Agent Development

  • Start simple, then specialize - Begin with a general description of your agent’s role, then add specificity as you discover what works and what doesn’t through testing.
  • Use examples in Edit Mode - Show your agent examples of ideal behavior. “When asked about X, respond like Y” is often more effective than abstract guidelines.
  • Define what NOT to do - Sometimes it’s easier to define boundaries. “Don’t make assumptions about user requirements” or “Never skip test coverage” can be powerful constraints.
  • Test edge cases in Run Mode - Don’t just test the happy path. Give your agent ambiguous requests, incomplete information, or conflicting requirements to see how it handles uncertainty.
  • Iterate based on real tasks - The best agent behavior emerges from real usage, not hypothetical scenarios. Use your agent for actual work and refine based on what you learn.
⌘I