Building 200 API Integrations with an Autonomous Agent

At Nango, we embarked on a journey to push the boundaries of autonomous agents in building API integrations. Our background agent, while not flawless, impressively generates around 200 integrations across five major APIs—Google Calendar, Drive, Sheets, HubSpot, and Slack—in just 15 minutes and under $20 in token costs. This task, which would traditionally take an engineer about a week, showcases the potential of autonomous coding agents.

## Our Setup

We designed a straightforward pipeline where we define interactions like creating calendar events or sending Slack messages. Our orchestrator sets up a workspace for each interaction, and an OpenCode agent independently builds and tests each interaction with access to a test account. Once completed, the orchestrator assembles these interactions into a cohesive Nango integration per API.

## Learning from Experience

Initially, we allowed agents to operate with minimal constraints to understand their capabilities. This approach revealed that agents could fill instructional gaps and make API calls with limited context, though they sometimes failed spectacularly. We learned not to trust agents blindly, as they often optimized for task completion over accuracy, leading to issues like reusing test data improperly or hallucinating commands.

## Addressing Agent Failures

To mitigate these issues, we implemented stricter directory permissions, improved instructions, and enforced checks to prevent agents from modifying test data. We also required agents to justify skipping SDK features, which often led them to use the intended abstractions.

## Post-Completion Verification

After an agent completes its task, our orchestrator verifies the results by re-running tests, checking for modifications, and inspecting internal dialogues for common issues. This rigorous verification ensures reliability despite the agent's autonomous nature.

## Debugging and Skills

When debugging, we found that tracing errors from their root cause rather than the final error message was more effective. Skills played a crucial role, encapsulating integration know-how and distributing it widely. This allowed us to iterate on best practices and publish updates easily.

## The Role of OpenCode

OpenCode proved to be an excellent fit for our project with its client-server architecture and robust SDK. Its open-source nature allowed us to inspect and adapt the platform as needed, ensuring our practices were transferable and not tied to a single tool.

## Conclusion

While agents aren't ready to autonomously ship every integration, with the right tools and verification, they can reliably handle significant portions of integration work. We're excited to offer our background agent to Nango customers and look forward to future advancements in AI models.