The Data Flywheel: Synthetic Data for Self-Improving Agents

The data flywheel in action

A data flywheel means that better agents produce better data, and better data produces better agents. Synthetic data plugs into this loop by providing high-quality, controllable examples. Teams can inject edge cases, rare workflows, or specific user behaviors that would otherwise take months to collect in production. Research on reinforcement learning for agents shows that synthetic data can improve sample efficiency by 20, 40 percent when blended with real experience. This is not theoretical. Stable Diffusion, language models, and robot learning pipelines already rely on synthetic augmentations to reach higher performance with fewer real samples.

Real numbers on quality vs real data

Quality matters more than volume. A 2023 study on synthetic trajectories for web agents found that a well-curated synthetic dataset outperformed 5x more raw real-world logs on downstream tasks. The synthetic set had higher coverage of edge cases and cleaner labels. Synthetic data also reduces annotation costs. Labeling a single realistic desktop interaction might cost $5, $15 depending on complexity. Synthetic generation can drop that to pennies per trajectory if the underlying simulation is stable. The tradeoff is realism. Synthetic outputs can still miss subtle context, nuance, or subtle bugs that only real interactions surface. The best approach is hybrid: use synthetic data to bootstrap learning, then fine-tune on real experience.

Techniques for realistic synthetic interaction data

• Replay real user sessions to build ground truth trajectories, then perturb inputs and outputs to create variations.
• Use procedural generation to build new workflows that would never happen in production.
• Apply feedback from real agents to correct synthetic outputs and close the quality gap.
• Mix synthetic examples with real logs during training to balance coverage and realism.

How Coasty fits into the data flywheel

Coasty runs computer use agents on real desktops and browsers. This means the synthetic datasets it produces reflect genuine interaction patterns, not idealized simulations. Coasty can capture workflows, error states, and browser behavior that are hard to reproduce in controlled environments. The resulting synthetic trajectories are labeled and ready for training or evaluation. Coasty operates as a custom, contact-led service. Teams work with the Coasty data team to define requirements, scope projects, and receive datasets tailored to their stacks and use cases. There is no self-serve portal and no fixed price list. Every engagement is scoped around the specific data needs of the project.