Rivet: Six Months Later

Six months ago, we introduced Rivet, an open-source solution that helps developers design, collaborate on, and validate enterprise-grade AI agents. It began life as an internal tool to visualize intricate AI logic loops but soon proved so useful that we added debuggability, shared code review, and testing features. We loved the prototype we built, but we realized quickly that the only way to make it as good as it could be would be to take it open-source.

Now that the tool’s been out in the wild for half a year, I wanted to share how it is making an impact with real development teams.

Navigating the “black box”

One of the most challenging aspects of agentic AI work, and a common obstacle facing development teams, is that LLMs can often feel so unpredictable. It can feel like working with a “black box,” where you try something and get a certain result, but can’t easily determine what, exactly, created that outcome. This can turn your design and testing into a trial-and-error process, slowing or stalling progress.

Justin Kwok, Software Engineer at Bento, a provider of in-product digital experiences that support customer activation and retention, saw Rivet’s impact when building the company’s first major AI release.

“Initially, every iteration we made seemed to result in unexpected outcomes, and we had no real way of knowing which tweaks improved things and which made things worse,” he says. “Rivet was a gamechanger for us. We used Trivet tests, a testing framework within the tool, to iterate systematically on prompts while minimizing regressions. It made the whole project so much easier.”

Managing complex requirements

Another challenge facing developers, particularly those creating enterprise agents, is managing highly complex and varied user requirements. When that complexity is combined with criticality – when the agent must deliver results in a high-stakes area, where any mistakes would carry significant consequence – the development task becomes significantly more difficult.

That was certainly the case for AutoRFP.ai, which sought to automate the process of completing requests for proposals (RFPs) and security questionnaires for technology companies. Needing to create an agent that could handle vast numbers of specific terms and requirements for each customer, CEO Jasper Cooper turned to Rivet to support his team’s development process.

“Rapid prototyping in Rivet allows us to move faster, seeing how different models and methodologies can improve the product’s performance,” he says. “The tool’s graph-based approach allows us to identify where performance is lacking quickly, even across extremely complex workflows. With Rivet as a core part of our stack, we’ve deployed the latest models and techniques faster than the competition, giving us a real edge and helping us grow extremely quickly.”

Extensibility and interoperability

The most valuable developer tools are those that connect to the technologies that matter to your project. The best ones come with plug-ins and support capabilities that let you work with best-of-breed solutions. That’s particularly important in AI development, when new LLMs and standards are constantly emerging. Since launch, Rivet has expanded greatly in terms of capabilities and support options. The tool now supports not just OpenAI and Anthropic, but also open-source LLMs like Mistral and Llama.

That was important to Wayne Chang, founder and CEO of Patented.ai, a company that applies the power of AI to solving customer intellectual property challenges. After exploring other AI development tools, his team chose Rivet because of its range and interoperability.

“Rivet is light-years ahead of other prompt IDEs,” he says. “Its intuitive interface and extensibility with plugins makes it clear it is built for even the most complex use cases like ours.”

Making prompt engineering easier

The value of “prompt engineering” – changing how prompts are phrased – has been one surprising aspect of generative AI. Subtle changes to prompts can massively change the efficacy of the reply. And because the skills to craft a good prompt are not necessarily the same as those needed to craft good software, non-developers can play a vital role in optimizing and configuring your agent.

For Adrien Maret, Staff Engineer at Didask, a provider of digital learning platforms, Rivet’s ability to split software engineering from prompt engineering was decisive. He found value in the tool’s ease of use and visual interface, which allows anyone to understand and work on prompt logic flows.

“Rivet allowed us to industrialize our use of LLM within our educational AI,” he says. “Prompt engineers are now able to iterate complex workflows on their own and then deploy them in production without requiring developers. This allows Didask to maintain our lead over the rest of the e-learning industry in terms of the use of AI.”

The best lies ahead

I love these stories because they show that Rivet is making a difference. It isn’t just a toy or interesting piece of technology to play around with… it’s a foundational solution that many companies, including Ironclad, are betting the future of their AI development on.

This tool might have started at Ironclad, but it is the Rivet GitHub community that has made it what it is today. I want to recognize their incredible effort and ingenuity. I am so thankful and impressed with their work and creativity.

I also want to thank a few of our partners who shared their use cases and helped spread the word. Niels Swimberghe of AssemblyAI shared an innovative podcast Q&A application. Pavel Duchovny of MondoDB posted about a new database vector search solution. Tim Kӧhler has used Rivet extensively in his AI Made Approachable channel, while Bryan Bischof of Hex gave it a memorable shout-out at the AI Engineering Summit.

I am blown away by the progress Rivet has made in just six months! I’m excited to see what the future holds. I hope you will consider using it in your own projects and share what you come up with!