Good Start Labs uses games as training environments to study which capabilities can be taught to AI models and whether those skills transfer to real‑world work. The company spun out of AI media and tools company Every in October 2025 and raised $3.6 million from General Catalyst, Inovia, Every, and angel investors.
Origin: a Diplomacy Twitch stream
Part of the company’s founding story traces back to a 2025 Twitch stream in which frontier models played the board game Diplomacy — a game that normally takes days or weeks to complete. Alex Duffy, co‑founder and CEO of Good Start Labs and formerly head of AI training at Every, observed that frontier models behaved differently in gaming contexts. He noted that the OpenAI o3 model often won by planning a future betrayal, while the Claude Opus 4 model refused to lie and was consequently ‘‘destroyed,’’ in his words.
Duffy argues these differences show that fine‑tuning on games can teach strategic thinking and other verifiable skills. He also wrote in a subsequent Every article that fine‑tuning on Diplomacy improved model performance on customer‑support and industrial operations benchmarks.
Case study: 1830 and financial research
Good Start Labs trained a 30B‑parameter model inside 1830: The Game of Railroads and Robber Barons, a strategy game described on Wikipedia as having the only luck element in determining initial play order. The game includes a stock‑market mechanic where players bid for shares in railroad companies to build a logistics network.
The team then tested the same model on financial‑research tasks: extracting information from a database about how the game had been played, placing it into an Excel‑like sheet, creating functions, and computing answers — steps that mirror typical finance workflows.
Published results compared a single‑turn question‑answering setup (the model sees a game state and picks the next move) with a multi‑turn terminal‑agent design that uses tools to explore the environment, plan strategy, and adapt in real time. Both training approaches improved their respective in‑game objectives, but only the terminal‑agent design improved performance on the Finance‑Agent benchmark. That suggests that training design — in particular goal‑directed, tool‑using, multi‑turn agents — matters for transferring skills to structurally similar external tasks.
Environment design and the role of the harness
Duffy emphasizes that not only the choice of game but also how the game is presented to the model affects what it learns. Good Start Labs can add expert models that provide denser, stepwise rewards, and the same game can be framed in different ways: a model looking at images learns different things than one reading natural text or one forced to frame everything in Python.
In the COS‑PLAY: Co‑Evolving LLM Decision and Skill Bank Agents for Long‑Horizon Tasks paper co‑authored by Duffy and Tyler Marques and researchers from several universities, the system gives a decision agent access to a ‘‘learnable skill bank’’ to guide action taking. A separate skill‑bank agent studies trajectories and adjusts the skill bank, which is then looped back for subsequent runs.
Duffy argues the harness — the environment and tool access — is particularly important. Even though more capable base models need less handholding, Good Start Labs treats the environment as curriculum, and the harness can enforce how a model solves a problem (for example, making it use code so results are auditable rather than relying on the model’s internal math).
Frontier models and measurement
Good Start Labs compares every new model; Duffy says newer, more capable models generally perform better in games but diverge on personality axes like betrayal, collaboration, and theory of mind. In Good Start Labs’ September 2026 rankings, Grok 4 Fast was the least likely to betray in Diplomacy, while Gemini 2.5 Pro proved less reliable in that respect.
Duffy also notes that models such as GPT‑6 Astra tend to report doing less chain‑of‑thought and jump to answers. The harness can require using explicit methods (e.g., code), which makes outputs verifiable even if a model can compute answers internally.
What Good Start Labs sells
The company describes its work as building ‘‘improvement loops’’ consisting of training systems, harnesses, and observability. Productized offerings are primarily twofold:
- Data: trajectories of agents playing games (what agents observed, decisions made, actions taken, and outcomes) and custom data from agents live in publishers’ games. Good Start Labs says any data sold to model developers is anonymized and stripped of personally identifiable information.
- Learning environments: full games where models can play end‑to‑end. The emphasis is not on winning but on teaching models to solve problems and use game engines as verifiable reward sources while performing potentially unexpected problem‑solving tasks.
Their main customers are frontier labs that use reinforcement‑learning data and environments to further train base models.
Can game skills transfer to real work?
Good Start Labs is also building a general model by unifying expert models trained on specific games into what Duffy calls ‘‘general game intelligence’’ that could be broadly applicable. Empirical evidence so far is mixed but promising: the 1830 finance experiment and Diplomacy training both showed improvements on analogous external tasks, and Duffy cites a Surge AI article and DeepSeek R1 results indicating post‑training improvements in office/work tasks.
Duffy summarizes the current state as showing that goal‑directed execution and reasoning transfer fairly clearly. However, he cautions that how broadly and reliably these game‑learned skills generalize across domains remains an open research question.
Conclusion
Good Start Labs applies a game‑based, verifiable training methodology to teach specific capabilities to AI models. Early experiments indicate that certain gameplay‑taught skills, especially those involving tool use and multi‑turn planning, can transfer to structurally similar real‑world tasks; but further work is needed to establish the breadth and robustness of such transfers.
