Mahesh Sathiamoorthy, who leads the AI startup Bespoke Labs, posted a thread on X arguing that curated RL environments, not additional model weight updates, have become the resource that determines whether an agent works in production. The claim reframes where builders should spend their compute and engineering time. It also describes, whether intentionally or not, the product category his own company sells.
His argument starts with a comparison to how deep learning matured. Teams settled on a repeatable recipe: curate a dataset, split it into training and test sets, and train a model that generalizes from one to the other. Sathiamoorthy’s point is that agent builders never converged on an equivalent recipe, because the data equivalent for agents, a scored, repeatable task environment in which an agent can act, has been scarce.
An RL environment, in his framing, is simply a simulated task paired with a scoring function: a coding repository with tests, a spreadsheet with a correct answer, a customer support scenario with a rubric. The “RL” label is a holdover from reinforcement learning research; the environments themselves are equally useful for supervised fine-tuning, prompt optimization, or plain evaluation. What matters is that the task and the score are defined once and reused, instead of an engineer eyeballing outputs and calling it done.
He breaks an agent into three tunable parts: the underlying language model, the system prompt, and the harness (the surrounding code that calls tools and manages state). Frontier labs mostly spend compute updating the first through reinforcement learning and supervised fine-tuning on environment-generated trajectories. Everyone else, he argues, is better served tuning the second and third, using optimizers such as GEPA to search for better prompts and emerging automation for harness design that he references as Meta-Harness work. Every one of those three levers still needs the same input: an environment that can score a trajectory as better or worse.
As evidence, Sathiamoorthy points to Snowflake’s chief executive comparing Zhipu’s GLM-5.2 against Anthropic’s Opus 4.7 using a set of 103 environments built around dbt, the data transformation tool. He notes that Bespoke Labs curated that dataset, which is worth flagging directly: the company making the case for environments as the industry’s next essential input is also the company selling them. That does not make the underlying argument wrong, but it means the comparison functions as both an evaluation and a sales demonstration.
The broader shift is visible beyond one company’s marketing. Frontier labs have converged toward similar benchmark scores on general capability, which pushes competitive advantage downstream into how well a specific agent performs on a specific customer’s tasks. A single set of fine-tuned weights belongs to one lab. A well built environment, task plus scorer, is licensable, resellable, and reusable across many customers’ harnesses and prompts, which is why a cohort of newer vendors alongside established data-labeling firms have repositioned around building and selling them rather than around labeling data for pretraining.
Teams currently benchmarking agents against public leaderboards should treat that practice as insufficient for procurement decisions. Building or buying a small set of domain-specific scored environments, even a few dozen tasks with clear pass criteria, will tell an operator more about whether an agent handles their actual workflow than any general benchmark will.
Posted by Mahesh Sathiamoorthy (on X) on August 4, 2026.
