How Agents 404 Devs Created a Prototype and Gained 120K Wishlists in Two Months
PlayHero Founder Anton Slashcev explains how Bezi helped his team prototype Agents 404 in one month, why core systems remain protected, and where human technical expertise is still essential.
Agents 404 is a physics-based co-op stealth comedy in which profoundly incompetent assassins infiltrate guarded locations, adopt ridiculous disguises, eliminate caricatured villains, and struggle to conceal the resulting evidence. Inspired by Hitman but designed around collaborative chaos, the game turns carefully planned assassinations into opportunities for physical comedy.
Developer PlayHero built the initial Agents 404 prototype in approximately one month, using systems carried forward from an earlier project and Bezi, an AI agent that works directly inside Unity. The resulting prototype helped the studio quickly validate the concept, assemble a trailer, and generate more than 60,000 Steam wishlists. Bezi assisted with features including disguises, camera behavior, effects, level construction, and the multiplayer ragdoll system that lets several players awkwardly carry the same body.
In this interview, PlayHero Founder Anton Slashcev discusses where AI assistance genuinely saves time, how PlayHero prevents technical debt from spreading, and why developers must remain responsible for every decision made during production.
Agents 404 is described as a physics-based co-op stealth comedy built around clumsy hitmen, chaotic plans, and things going wrong in funny ways. What was the original creative spark for the game that led you to something like this?
Anton Slashcev, PlayHero Founder: The core idea was inspired by the Hitman series, but reimagined as a co-op experience that is more playful, chaotic, and less demanding or overly serious.
What kind of projects had you and your team worked on previously?
Anton Slashcev: I have an extensive background in mobile game development. I worked as a producer at Playrix and Belka Games on billion-dollar games. I also founded my own studio, Unlock Games, where we released several hybrid-casual hits, including Detective Masters and Become a Queen.
You have posted about using Bezi to prototype Agents 404 and other PlayHero projects. What does AI-assisted prototyping actually look like? Can you walk us through one specific feature or mechanic in Agents 404 that started as an AI-assisted prototype?
Anton Slashcev: One feature that we developed with extensive help from Bezi was the ragdoll system that allows several players to carry a body together.
We maintain documentation for every feature and major aspect of the game, which was also created with Bezi’s assistance. When we start working on a new feature, we first provide all the relevant context. Then we describe the requirements from a game designer’s perspective. If the feature is complex, we also specify technical constraints and architectural patterns to follow.
Next, we ask Bezi to raise clarifying questions, suggest possible improvements, and prepare an implementation plan. For complex features, our lead developer reviews the plan to make sure the proposed approach is sound. Once the plan is approved, the agent writes the code and implements the feature in the game.
When you are testing a new idea, how do you decide what to ask Bezi to build versus what should be handled directly by a designer, programmer, or artist?
Anton Slashcev: It depends on the task. First, it is important to distinguish between writing code and working with assets.
When it comes to code, Bezi handles most features that do not affect the core architecture. The same applies when there is already an existing implementation pattern that it can follow. If the task is new, complex, or architecture-sensitive, our lead developer will usually handle it, potentially with Bezi’s assistance.
Asset-related tasks are still usually handled by a game designer or artist. Bezi helps them work faster by answering questions about why something is not working or explaining how to configure a particular component.
Do you feel like it required significant technical expertise to actually be useful and actionable?
Anton Slashcev: In my opinion, significant technical expertise is not required if you are building a prototype. However, if you want to create a scalable, production-ready system, then yes, a certain level of technical expertise is necessary.
You’ve talked publicly about team members beyond programmers doing more technical work with Bezi. How has that changed collaboration between designers, artists, and engineers at PlayHero?
Anton Slashcev: The main benefit is that it reduces the workload on developers. Before artists or game designers approach a developer with a question or a request for functionality needed for visual or cosmetic purposes, such as spawning VFX or triggering an animation, they first try to solve it with Bezi. If they cannot, they then go to a developer.
Agents 404 depends heavily on physics, co-op interactions, stealth, and comedy. Those systems can be fragile because small bugs or tuning changes can completely alter the feel. How do you use Bezi while still maintaining control over design intent and game feel?
Anton Slashcev: Physics, co-op interactions, and stealth are indeed complex and fragile systems. This is exactly where technical expertise and a deep understanding of the project are still essential. You need to understand how the system works and what changes the AI is making.
Critical or complex systems should never be black boxes to the team. Someone must understand them thoroughly. For isolated modules or one-off features, you do not necessarily need the same level of technical involvement.
What guardrails do you put around AI-assisted development? For example, how do you define coding standards, module boundaries, review steps, documentation, or areas where Bezi should not make changes?
Anton Slashcev: We have several systems, such as our generalized NPC behavior simulation, that AI is not allowed to modify when implementing features. It can only interact with these systems through their existing interfaces. This prevents unintended changes to core logic.
When AI-generated code affects a complex or sensitive part of the game, our lead developer always reviews it. We also try to provide clear architectural patterns for the AI to follow.
A lot of developers worry that AI-generated code can look correct while hiding technical debt, bugs, or performance problems. How do you review, test, and validate AI-assisted work before it becomes part of the real project?
Anton Slashcev: This is precisely why we isolate certain parts of our systems and prevent AI from modifying them directly.
AI-generated code can create technical debt and may not always be ideal. That can be acceptable when it happens inside an isolated module. It becomes a serious problem when it spreads throughout the entire system.
At the same time, when the AI is given a clear coding pattern defined by the lead developer, it can often produce solid, functional code. The real skill lies in recognizing when the AI is proposing a good solution and when it is proposing a bad one. That is exactly where technical expertise remains essential.
The public response to Agents 404 seems to show that rapid prototyping can still lead to something with strong market appeal, including major wishlist momentum. How do you balance moving quickly with making sure the idea is actually fun, polished, and commercially understandable?
Anton Slashcev: The initial Agents 404 prototype took about one month to develop. However, it was built on top of systems created for our previous prototype, Titan Hunters, which took several months to develop.
During the early validation stage, speed is more important than quality. The goal is to determine whether the core idea works and whether players respond to it. Once the concept has been validated, the focus naturally shifts toward improving quality, polish, and scalability.
For 80 Level’s audience of developers and technical artists, what have you learned about using AI assistants responsibly in game development? Where are they genuinely useful today, and where do you think people still overestimate what they can do?
Anton Slashcev: The most important thing to remember is that using AI does not remove responsibility from developers or eliminate the need for human decision-making.
AI is an executor. Sometimes it can also act as an adviser or consultant. However, the decisions themselves must still be considered and made by people. AI is still far from being consistently good at decision-making, and that is the crucial limitation.
Decisions about project architecture, visual style, game design, UX, and other fundamental aspects have not disappeared, and they cannot simply be delegated to AI. What AI can do is dramatically accelerate the implementation of those decisions. It can also help teams understand the project context faster and make better-informed choices.
In short, I believe in AI-assisted development, rather than generative AI or an AI-first approach
