Zencoder, an artificial intelligence startup focused on software development, today announced the launch of a suite of products developed using generative AI agents that are designed to help software developers work faster and build better applications.

Founded in 2023 by Chief Executive Andrew Filev, Zencoder offers AI agents that integrate into development workflows to provide intelligent code generation, code repair, test generation, optimization and documentation.

Filev told SiliconANGLE in an interview that Zencoder differentiates itself from other AI-powered coding assistants on the market by bringing AI agents to developers. The AI agents are capable of multistep coding capabilities and self-repair, which sets the company apart from the competition.

“The use of the original coding assistant by a lot of developers is compared to StackOverflow on steroids,” said Filev. He added that for the most part assistants provide code-snippet suggestions and knowledge. These use cases, although extremely useful, have left people feeling oversold.

To build on the AI agents as companions, Filev said that his company invested heavily in a product called Repo Grokking, a tool that allows the company’s agents to deeply analyze a company’s codebase to understand the structure, understand the structure, identify patterns and the logic behind code. It allows the agents to fine-tune themselves to a company’s codebase to provide more relevant suggestions and better responses to queries and prompts.

“So, you can ask within the chat for it to solve simple bugs for you, for example, or implement simple new features, and it will plan its work,” said Filev. “It will try to execute that work and try to validate that work before it presents it to you.”

This highlights the second innovation of the agent workflow that Filev says sets Zencoder apart, what he calls agentic repair. When an AI agent produces code, it will automatically analyze and refine the generated code in an attempt to fix any potential bugs that might have been produced or further improve its output. It can do this in the context of the codebase using its awareness of how the code interacts with other parts of the application and will even run tests to make certain it will compile without issues.

“We invested a lot in this component that analyzes the code that AI generates and tries to fix it,” Filey said. “When it does, it uses both AI and non-AI techniques, which you can call assembling. A lot of our team have backgrounds both in AI and in things like compilers and code parsers, so we understand code really well and the tooling of how you work with the code.”

When an AI agent generates code, he explained, it can use syntax analysis to quickly discover mistakes and run it again with an AI to find and fix mistakes before it is offered back to the user. The repair is a multistep process designed to help mitigate potential hallucinations or errors that could happen during the AI code generation process.

Zencoder supports major programming languages including Java, JavaScript, TypeScript, Python and C#. It can also integrate with popular editors such as Visual Studio Code and JetBrains.

Filev said the company plans to launch custom agents in the future, allowing users to train an agent to automate repetitive tasks that happen during the development process. For example, a development team may want to migrate from one version of a library to another, or from one infrastructure to another, making a lot of little changes all across the codebase.

That would normally be done with a series of different macros and then some manual changes for everything that was missed. However, with Repo Grokking and an AI agent, much of this work could be completed by asking what changes needed to be made and the best scripting method to do it.

“It’s not an agent that pretends that it’s going to do the whole work for you,” said Filev. “It’s your companion, kind of your little AI intern, that can help you recreate what you want to do.”

 Image: Zencoder