AI Agents for App Maintenance | logicstar

LogicStar Enters the AI Agent Arena with a Focus on Software Maintenance

A newly established Swiss startup, LogicStar, is making its entry into the burgeoning field of AI agents. Founded in the summer of 2024, the company has successfully secured $3 million in pre-seed funding.

This investment will be utilized to develop tools specifically for developers, enabling the autonomous maintenance of software applications. This represents a departure from the more common application of AI agents, which typically centers around collaborative code development.

A Complementary Approach to Code Development

LogicStar’s CEO and co-founder, Boris Paskalev, envisions a synergistic relationship between their AI agents and existing code development agents. He suggests potential partnerships with platforms like Cognition AI’s Devin, creating mutually beneficial business outcomes.

Maintaining code fidelity is a critical challenge for both AI agents and human developers. LogicStar aims to streamline the development process by automatically identifying and rectifying bugs within deployed code.

Addressing Limitations in Current AI Models

Currently, Paskalev notes that even the most advanced models and agents struggle to resolve a significant portion of the bugs they encounter. This observation prompted the team to identify an opportunity for a startup dedicated to enhancing bug resolution rates and reducing the burden of tedious app maintenance.

To achieve this, LogicStar is leveraging the power of large language models (LLMs), including OpenAI’s GPT and DeepSeek from China. Their platform adopts a model-agnostic approach.

A Model-Agnostic Platform for Optimal Performance

This flexibility allows LogicStar to utilize various LLMs, maximizing the utility of its AI agents by selecting the foundational model best suited for addressing specific code issues. The goal is to optimize performance across a diverse range of programming challenges.

Leveraging Expertise and Past Success

Paskalev emphasizes that the founding team possesses both the technical expertise and specialized knowledge necessary to build a platform capable of resolving complex programming problems that may challenge LLMs operating independently.

Furthermore, the team has a proven track record of entrepreneurial success. Paskalev previously sold his code review startup, DeepCode, to the cybersecurity firm Snyk in September 2020.

Focusing on Business Value Extraction

Initially, the team considered developing their own large language model for code. However, they quickly realized this would likely become a standardized commodity. Instead, they shifted their focus to maximizing business value from existing LLMs and AI agents.

“Now we’re building assuming all those large language models are there,” Paskalev explained to TechCrunch. “Assuming there’s some actually decent [AI] agents for code, how do we extract the maximum business value from them?”

Grounding and Verification for Reliable Results

The core idea stems from the team’s deep understanding of software application analysis. By combining this knowledge with large language models, and focusing on grounding and verifying the suggestions made by these models and AI agents, LogicStar aims to deliver reliable and effective solutions.

Test-Driven Development: A Practical Approach

What does this methodology entail in a real-world context? According to Paskalev, LogicStar conducts a detailed analysis of each application where its technology is implemented, utilizing established “classical computer science methods.” This process constructs a comprehensive “knowledge base.”

This knowledge base provides the AI agent with a complete understanding of the software’s inputs and outputs, the relationships between variables and functions, and all other relevant connections and dependencies.

When presented with a bug, the AI agent can then pinpoint the affected areas of the application. This allows LogicStar to focus its simulation efforts on only the necessary functions, enabling the testing of numerous potential solutions.

Paskalev explains that this “minimized execution environment” empowers the AI agent to execute “thousands” of tests, aiming to replicate the bug and identify a “failing test.” Through this “test-driven development” process, a reliable fix is ultimately achieved.

He clarifies that the actual code corrections originate from Large Language Models (LLMs). However, LogicStar’s platform, with its “very fast execution environment,” allows its AI agents to operate at scale, effectively filtering results and providing users with access to the most effective solutions offered by LLMs.

“LLMs demonstrate considerable promise in prototyping and experimentation, but they are currently not well-suited for full-scale [code] production or commercial applications. We believe significant advancements are still needed in this area, and our platform addresses this challenge,” he stated. “Our ability to harness the capabilities of these models allows us to extract commercial value and free up developers to concentrate on more critical tasks.”

LogicStar’s initial focus will be on enterprise clients. Its “silicon agents” are designed to work alongside existing corporate development teams, offering a cost-effective alternative to hiring additional human developers, and handling routine application maintenance tasks.

While the startup promotes a “fully autonomous” application maintenance capability, Paskalev confirms that human developers will have the option to review and oversee the fixes proposed by the AI agents. Building trust is therefore a primary consideration.

“The accuracy typically achieved by a human developer falls within the 80 to 90% range. Our objective [for our AI agents] is to reach that same level of precision,” he adds.

LogicStar is currently in its early stages, with an alpha version of its technology undergoing testing with a select group of undisclosed “design partners.” Currently, the technology supports Python, with plans to expand to TypeScript, JavaScript, and Java “coming soon.”

“The primary goal [of the pre-seed funding] is to demonstrate the technology’s effectiveness with our design partners, initially focusing on Python,” Paskalev explains. “We have already invested a year in development and see significant opportunities for expansion. Therefore, we are prioritizing a focused approach to showcase the value proposition in a single use case.”

The startup’s pre-seed funding round was led by European VC firm Northzone, with participation from angel investors associated with DeepMind, Fleet, Sequoia scouts, Snyk, and Spotify.

In a released statement, Michiel Kotting, a partner at Northzone, commented: “AI-driven code generation is still in its nascent stages, yet the productivity improvements we are already witnessing are transformative. The potential of this technology to streamline development, lower costs, and accelerate innovation is substantial. The team’s extensive technical expertise and proven history position them to deliver tangible, impactful outcomes. The future of software development is evolving, and LogicStar is poised to play a vital role in software maintenance.”

LogicStar is currently maintaining a waiting list for potential customers interested in gaining early access. A beta release is anticipated later this year.

Stay informed with TechCrunch’s AI newsletter! Subscribe here to receive it in your inbox every Wednesday.