Codegen raises new money to automate software engineering tasks

Jay Hack, an AI researcher with a background in natural language processing and computer vision, came to the realization several years ago that large language models (LLMs) – think OpenAI’s GPT-4 or ChatGPT – have the potential to make developers more productive through natural language requests to translate into code.

After working at Palantir as a machine learning engineer and building and selling Mira, an AI-powered cosmetics retail startup, Hack began experimenting with LLMs to perform pull requests — the process of merging new code changes with the main project repositories. With the help of a small team, Hack slowly expanded these experiments into a platform, Codegen, that seeks to automate as many mundane, repetitive software engineering tasks as possible using LLMs.

“Codegen automates the menial work of software engineering by enabling AI agents to submit code,” Hack told TechCrunch in an email interview. “The platform enables companies to move significantly faster and eliminates the costs of technology debt and maintenance, allowing companies to focus on product innovation.”

So you might wonder: what differentiates Codegen from code-generating AI like GitHub Copilot, Amazon CodeWhisperer, and the Salesforce model with which Codegen shares a name? First, the challenges that Codegen addresses, says Hack. While Copilot, CodeWhisperer, and others focus on code autocompletion, Codegen deals with “codebase-wide” issues like major migrations and refactoring (that is, refactoring an app’s code without changing its functionality) .

“Codegen uses a multi-agent system to generate complex code,” Hack explains. “This involves orchestrating a swarm of agents to collectively break down and solve large tasks. Many LLMs deliberate effectively and build on each other’s work, [which] produces significantly better results.”

READ MORE  White House wants to reduce software costs for government agencies
Codegen gif

Codegen gif

Image credits: Codegen

Codegen’s core product is a cloud and on-premises tool that connects to codebases and project management boards, such as Jira and Linear, and automatically generates pull requests to handle support tickets. The platform can even set up some of the necessary code infrastructure and logging, Hack says, although it wasn’t clear to this reporter what Hack meant by “infrastructure.”

“Unlike other solutions, Codegen offers a higher level of automation when performing entire tasks on behalf of developers,” said Hack. “We scrape a company’s backlog, find the solvable tickets, and then set up an army of agents to find the relevant code and produce a pull request.”

Now Codegen promises a lot, as even today’s best AI models make big mistakes. For example, it is well known that generative coding tools can introduce unsafe code, with a Stanford study finding that software engineers who use code-generating AI are more likely to introduce security vulnerabilities in the apps they develop.

Hack says that Codegen, in turn, is trying to find “the right balance” between human oversight and best practices around monitoring LLM-generated code.

This is important work, and the entire development ecosystem would benefit from a better understanding of how to evaluate and verify LLM output,” said Hack. “Significant progress will need to be made before widespread developer confidence in generalized , automated code generation systems. .”

Investors seem to think Codegen is heading into the future, for what it’s worth.

The company announced this week that it has closed a $16 million seed round led by Thrive Capital with participation from angel investors including Quora CEO Adam D’Angelo and Instagram co-founder Mike Krieger. The tranche brings Codegen’s total raised to $16.2 million and values ​​the startup at $60 million, Hack claims.

READ MORE  This cloud software stock will almost double by 2023: is it too late to buy?

Thrive’s Philip Clark said via email: “In 2023, most developers still spend an unreasonable portion of their time writing code for low-level tasks like migrations, refactorings, integrations, and bug fixes. Companies like Codegen are leveraging of LLMs to build AI agents that free engineers from this tedious work. Developers will soon be able to offload jobs to agents so they can stop worrying about software toil and continue focusing on creating new products.”

San Francisco-based Codegen doesn’t have any paying customers yet — it’s currently incubating the platform with two “large-scale” business partners. But Hack expects growth in the coming year.

“We are raising significant capital because the opportunity to create such a substantial and ambitious product has only recently arisen, and we want to sprint to market with full steam,” he said, adding that Codegen plans to build out its workforce expanding from six employees to 10 by the end of the year. “The funds will be used to scale our workforce and support our infrastructure.”