The Antithesis Approach to Breaking AI-Generated Code

Over the last two years, AI leaders have been predicting the downfall of the traditional software industry—or SaaSpocalypse—as AI writes most, if not all, of the code for engineering teams today. Seldom can one find programming that isn’t handled by the likes of Claude Code and Codex. 

But knowing where the project breaks, how it breaks, and what happens after is now more important than ever in the agentic era. 

It was almost as if Antithesis was waiting for this moment. 

Founded by Will Wilson and Dave Scherer eight years ago, the autonomous software testing company operated in stealth mode for an unusually long time—until 2024. Wilson and Scherer were also the creators of the high-speed NoSQL database FoundationDB, which was acquired by Apple in 2015. 

Antithesis was born out of Wilson and Scherer wanting to invent a custom testing method to ensure FoundationDB was rock-solid and fault-tolerant under chaotic real-world conditions.

The company calls it ‘deterministic simulation testing’, running a full software stack inside a custom hypervisor. 

By controlling sources of non-determinism—like time, thread scheduling, and network latency—software becomes a pure state machine. If a bug occurs during a stress test, the environment can replay the exact sequence of events with 100% fidelity.

This is the bet behind the nearly $182 million the company has raised, including a $105 million Series A led by Jane Street last December. 

“Producing code has suddenly become very cheap, but that has exposed how much work we traditionally had to do after writing code before it reached users,” Akshay Shah, Field CTO of Antithesis, says in an interaction with AIM.

The traditional model of writing code, peer reviews, testing, and spending a lot of time ensuring it doesn’t break is counterproductive today. “That model held together because the rate at which we could write code matched the rate at which we could analyse, review, and verify it,” says Shah. “But nobody has the time for that level of manual review and testing.”

Antithesis’s products operate as an autonomous testing environment. For example, its ‘Multiverse Debugger’ is an interactive tool that provides time-travel debugging. When a fault is triggered during a simulation, developers can rewind execution, inspect system state, and perform root-cause analysis to isolate the failure event by event.

To fit modern developer workflows, the company provides native agent skills and APIs tailored for AI tools like Claude Code and OpenAI Codex. These integrations allow AI coding agents to write setup manifests, run stress tests in Antithesis, receive structured failure reports, and patch their own code in a closed loop.   

Antithesis also maintains two open-source frameworks: Hegel, a multi-language property-based testing engine for application logic, and Bombadil, a property-based testing tool for web and terminal user interfaces. 

Antithesis Caught Bugs That Claude Fable Couldn’t

Recently, the company announced an agent-based mutation-testing workflow that deliberately injects bugs to determine whether AI-generated tests can actually catch them. 

In its rqlite experiment, the agent tested 13 safety properties and successfully falsified 11 after 19 mutant injections and 46 runs; the remaining two required further testing or expert reconsideration of the properties

While developing the workflow, Antithesis says its tests also uncovered three previously unknown bugs in rqlite, two involving safety or correctness and one involving liveness. The bugs appeared to have escaped a full code scan using Claude Fable, highlighting the difference between an agent reading code for potential problems and an environment that subjects the system to the conditions under which those problems emerge.

“What Antithesis does is take these complex systems, place them into a specially designed environment, and stress-test them far more rigorously than they would ever be tested in real life,” says Shah. “That allows us to accelerate the hardening process—turning what normally takes years of production outages into a few minutes or hours.”

If the code breaks, the agent gets a detailed, reproducible bug report that it can fix in its next iteration. 

Antithesis is used by companies including Jane Street, MongoDB, Ramp, and Palantir, among others. 

The company says MongoDB has used its tools since 2021, expanded it into eight network topologies, found 100+ critical bugs, reduced mean time to resolution by 47%, and increased its annual investment in its testing by more than 3x.

How Do Deterministic Simulation Environments Work? 

In computer science, the Halting Problem asks whether a computer program will finish running or run forever in an infinite loop based on a specific input. The problem is undecidable, but it helps Antithesis determine whether a set of simulations is good enough. 

Shah explains that the tools measure how much new ground each additional unit of simulation time covers and share a ‘utilisation graph’ with users as an output. 

“When you see that graph climbing sharply, you know you should keep simulating. Eventually, the graph flattens out, indicating you have reached a point of diminishing returns,” he says. 

At that point, a flat graph could mean either the system has been thoroughly exercised, or you haven’t tested the right properties yet.

“For instance, if you are testing payment settlement software, you might exhaustively exercise two-party transactions, simulating server crashes, network drops, and phone disconnects, and confirm it works flawlessly,” explains Shah. “But if you completely forgot to test a three-way bill-splitting feature, that is an entire block of code that remains unexercised.” 

This is also where test quality depends on what engineers ask the system to exercise. Antithesis needs a workload that drives the software; engineers can then add properties describing behaviours that must always hold. 

The platform can explore an enormous state space, but it cannot infer every requirement the system must satisfy. “Antithesis is like a power saw. It is incredibly useful, but it doesn't build the furniture by itself.” 

Antithesis runs customer software as standard containerised workloads (such as Docker or Kubernetes clusters) inside a specialised, proprietary hypervisor. This hypervisor sits beneath the guest operating systems, creating an isolated sandbox that completely abstracts the underlying host hardware.

For the simulation to be truly deterministic, every source of unpredictability is intercepted and controlled. The hypervisor virtualises four main sources—system time, entropy, thread scheduling, and network and input/output operations.

Essentially, it continuously generates different inputs, faults, and execution timelines while checking whether the system violates properties engineers have defined. Once Antithesis finds a failure, reproducibility kicks in. 

Antithesis runs the software in a deterministic environment, allowing a developer to return to the exact point where a failure occurred and replay the sequence of events that produced it.

“In areas where correctness isn't critical, many teams have simply relaxed their standards and accepted higher bug rates,” says Shah. “But in domains where correctness is paramount—payments, databases, infrastructure, blockchains, and financial trading—teams are actively struggling for answers.” 

Onwards and Upwards

Going back to the company’s roots in 2018, Shah says the name emerged from an inside joke. “From the beginning, the company took a contrarian approach—whenever the rest of the industry did things one way, Antithesis did the opposite.”

When the company started, few people thought testing distributed systems would become a venture-scale business. During its Series A funding, the company revealed that its revenue had increased by 12x over the preceding two years. 

“It turned out they were wrong, and the challenges of testing distributed systems turned out to be the exact same challenges involved in testing LLM-authored software.”