Request an evaluation(opens in a new tab)
Request an evaluation(opens in a new tab)

Transform your SDLC with agents.

From copilots to a self-driving SDLC,with the team behind the first AI for code.

Backed by frontier coding agent research with ETH Zurich.

Request an evaluation(opens in a new tab)Find your starting point
  1. 01Same ticket, same start. Today it waits between every step, the longest for review.
  2. 02With agents, each step that holds is automated. Ticket 1 ships before today’s is even reviewed.
  3. 03Same time: one ticket ships today. With agents, you ship the next ones too.
Illustration: today a ticket moves through plan, code, review, test and merge with long waits in between, the longest for review. With agents, each step that holds is automated, the waits go, and in the same time the next tickets ship too.
  • First AI for code

    We built DeepCode, the first commercial AI for code, in 2019. It’s now Snyk Code.

  • $100M+ ARR

    for Snyk Code, scaled by our CEO.

  • 300+ publications

    Frontier coding agent research with ETH Zurich, at NeurIPS, ICML and COLM.

  • 2+ years

    of building and operating coding agents in production.

  • $6M raised

    From Northzone, BEK Ventures and angels.

What our customers’ CTOs say

  • LogicStar catches those bugs before they get to production.

    Federico MuñozCTO, Sequence

  • The agent’s quality right now is much better than when we started.

    Waldemar HummerCTO, LocalStack

  • Something that changes what I plan to do for the next few weeks provides value.

    Johan TibellCTO, Benetics

Trusted by

  • Google DeepMind
  • Snyk
  • Spotify
  • Fleet
  • Sequoia

Where to start

We start where your team is.

Three stages of the journey. Find yours, and we’ll take you to the next one.

Mark what sounds like your team

What we’ve built

Done with customers, measured on their code.

Three customersMeasured on their own repositoriesWhat happened after merge

A midsize SaaS company

Human review safely removed on 50% of PRs

865 PRs analyzed. 82 review misses found.

After

Before

  • 1 in 10
    PRs shipped a severe bug to production
  • 1 in 5
    PRs the agent graded “perfect” still introduced a bug
  • 78%
    of PRs human review was redundant, yet still required. On the rest, it prevented 8 functional issues from reaching production

What we built

An auto-merge classifier for their review flow

A large SaaS company

Their review agent measured on actual performance, 4 failure modes identified for remediation

1 in 5 severe bugs escaped both AI and human review. 51 severe bugs observed.

Where each severe bug was caught

What we found

  • 58%
    caught by the agent before merge
  • 22%
    missed by the agent, caught by a human before merge
  • 20%
    missed by agent and human, and deployed

What we built

Measurement of their code review agent, on 51 severe bugs

A European engineering organisation

Software engineering tribal knowledge turned into quality standards they can automatically enforce

15k review comments across 5 projects turned into 117 transferable rules.

From review comments to rules

What the rules caught

  • Over 300
    later-fixed issues detected on held-out PRs
  • 31%
    of findings used rules transferred across projects

What we built

Quality rules mined from their own review comments

Each result comes from an engagement, measured on the customer’s own repositories and what happened after merge.

Request an evaluation(opens in a new tab)See how we work

Team

Built by people who have done this before.

Researchers and engineers who have built and run coding agents in production for over two years, from ETH Zurich and the team behind DeepCode (now Snyk Code).

Founders

  • Boris Paskalev

    Boris Paskalev

    Co-Founder and CEO

    Serial entrepreneur with a CS Master's from MIT. Co-founded DeepCode (acquired by Snyk) and led global teams of 100+ across SaaS and AI ventures.

  • Mark Müller

    Mark Müller

    Co-Founder and CTO

    PhD from ETH Zurich in code analysis for neural networks. 30+ research papers, 1300+ citations, with industrial work at Porsche, Mercedes, and Airbus.

  • Martin Vechev

    Martin Vechev

    Co-Founder and Advisor

    Martin is a full professor at ETH Zurich and one of Europe’s leading AI researchers, with 200+ publications in AI, networking, programming paradigms and others. He has co-founded five deep tech startups (including two exits) and launched INSAIT, raising over $280M to drive AI innovation.

And the team

Meet the whole team

Built on research

We know when agents work.

We publish benchmarks for AI coding agents, at NeurIPS, ICML and COLM. The same methods measure the agents you run.

8 papers2024–2026NeurIPSICMLICLRCOLM

  1. Write Test SuitesarXiv 2026

    56%

    of specifications tested by the best agent’s test suites

    TestSuiteBench: Coding Agents Struggle to Generate Strong Test Suites from Specifications

    Paper : TestSuiteBench: Coding Agents Struggle to Generate Strong Test Suites from Specifications (opens in a new tab)Leaderboard : TestSuiteBench: Coding Agents Struggle to Generate Strong Test Suites from Specifications (opens in a new tab)Read the post : TestSuiteBench: Coding Agents Struggle to Generate Strong Test Suites from Specifications

  2. Knowing when to actCOLM 2026

    35–65%

    of stale bug reports triggered changes to correct code

    Coding Agents Don’t Know When to Act

    Paper : Coding Agents Don’t Know When to Act (opens in a new tab)

  3. Refactor CodebasesICML 2026

    7.9%

    of agent-proposed refactors match human choices

    CodeTaste: Can LLMs Generate Human-Level Code Refactorings?

    Paper : CodeTaste: Can LLMs Generate Human-Level Code Refactorings? (opens in a new tab)

  4. Context filesMemAgents @ ICLR 2026 (Oral)

    +20%

    inference cost from context files like AGENTS.md, with no gain in success

    Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?

    Paper : Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? (opens in a new tab)

  5. SecurityICML 2025 (Spotlight)

    ~50%

    of functionally correct AI-generated backends are vulnerable to real exploits

    BaxBench: Can LLMs Generate Correct and Secure Backends?

    Paper : BaxBench: Can LLMs Generate Correct and Secure Backends? (opens in a new tab)Leaderboard : BaxBench: Can LLMs Generate Correct and Secure Backends? (opens in a new tab)Read the post : BaxBench: Can LLMs Generate Correct and Secure Backends?

  6. Benchmark GenerationICML 2025

    What works best depends on your task distribution.

    Automated Benchmark Generation for Repository-Level Coding Tasks

    Paper : Automated Benchmark Generation for Repository-Level Coding Tasks (opens in a new tab)

  7. Reproduce IssuesNeurIPS 2024

    2×

    better fix correctness when filtered by reproducing tests

    SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents

    Paper : SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents (opens in a new tab)Leaderboard : SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents (opens in a new tab)Read the post : SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents

How we work

Start small. Prove it. Then scale.

Embedded in your team. Where someone owns AI in engineering, we work alongside them; where nobody does yet, we take that role with you.

  1. Start small

    We start with the low-hanging fruit that delivers the most value. In the first week we measure where your engineering time and bugs actually go, on your own history. Then we agree with you the one place where agents will pay off first, and start there.

    You getA first agent running on your code, measured, and yours to keep.

  2. Prove it

    Every result measured on what actually happened in your repos: bugs caught, bugs escaped, review saved, cost.

    You getBefore-and-after numbers you can defend.

  3. Scale

    Hand more to agents, one workflow at a time, measured as models, prompts and tools change. Your engineers learn to run it with us, so it keeps working when we leave.

    You getMore of your engineers’ time on what matters, and agents taking the rest.

A few teams at a time

We choose our teams and take on only a few at once, so we go deep instead of spreading thin. The engineers who start with you stay with you.

Your code stays yours

Access is read-only by default, and nothing we or our agents produce reaches your code without your approval.

For your AI lead

The technical depth.

The software factory

How a software factory works, part by part.

A technical walkthrough for whoever runs AI in your engineering: how a ticket moves through a factory, what it is built from, what it can take today, and where to start.

IllustrationRead the walkthrough

The platform

We built a platform to measure our customers’ code review agents.

It keeps running for teams that want ongoing monitoring: results, real-world outcomes, and what changes after a new model, prompt or tool.

IllustrationEvery agent you run, measured continuously.

FAQ

What CTOs ask us first.

Getting started

What do you need from us to start?

You, involved in person, and read-only access to the repositories the work touches. Once we build or change agents, some time from one of your engineers. Nothing we or our agents produce reaches your code without your approval.

How soon do we see results?

The first piece of work takes a few weeks and ends with an agent running on your own code, measured and yours to keep, so you judge it on your numbers, not on our slides.

Which agents and tools do you work with?

Any agent whose work lands in your repositories: Claude Code, Codex, Copilot, Cursor, bought tools or your own, on GitHub, GitLab or wherever your code lives. We measure and improve what you run; we don’t replace it, and you don’t change your workflow. No agents yet? We set up the first one with you.

Your code and your people

Where does it run?

We’re flexible on where it runs: bring your own LLM endpoint or use ours. We start read-only, under NDA, and never train on your code. The Trust page covers the rest of what your security team will ask.

Do you measure our developers?

No. We measure agents, not people. Humans decide where agents may merge on their own, and nobody’s personal stats end up on a dashboard.

How do we get the team on board?

With evidence from your own SDLC. It gives teams a reason to adopt instead of a mandate to comply with.

Is the goal to remove humans from the loop?

Only where the evidence says it’s safe, your regulatory environment allows it and you want it. Human review comes down step by step, each step backed by evidence from your own history, and your people keep control throughout.

Before you commit

Where do the numbers come from?

From your own history. Every result traces back to a real PR, fix or incident in your repositories, not to opinions or an LLM grading its own work. When an agent missed something, reality already wrote it down.

Won’t the model providers build this themselves?

Measurement only works when it’s independent of the thing measured. We don’t sell you a model, so we have no stake in which one wins: your agents are judged on your outcomes, across whatever models you run, and the numbers stay honest when you switch.

What happens if we stop?

You keep everything: the agents and workflows we built with you, the labels, the prompts, the reports. Exportable by design, so leaving is always cheap. We’d rather earn the next piece of work than lock the door.

Transform your SDLC with agents.

Tell us where your team is with AI.We’ll help you reach the next level.

Request a 30-minute evaluation(opens in a new tab)transform@logicstar.ai