Salesforce lifts an AI agent's browser task completion from 43.5% to 93% without touching the model

Salesforce researchers took an AI agent from finishing 43.5% of browser tasks to 93% without touching the underlying model, according to a VentureBeat report. The result points to something agent builders increasingly argue: real-world agent performance is not decided by model capability alone.

The report frames the gain as an improvement made without touching the model — no retraining, no fine-tuning, and no swap to a stronger checkpoint. Instead, the lift came from the engineering around the model, which shifts the optimisation question from which model is best to how well the workflow is built.

Browser tasks are a demanding test because they are multi-step and long-horizon: open a page, judge its state, act, then confirm the result. Any single failure can break the whole run, so a 43.5% completion rate means more than half of attempts did not reach the end.

Closing that gap usually comes from redesigning the failure points — clearer task decomposition, more reliable judgements about page state, retry and recovery after errors, and validation of intermediate steps. The report attributes the improvement to that layer rather than to any change in the model's own abilities.

The signal worth taking from this is that as frontier model capabilities converge, agent products will compete on engineering detail. Teams that decompose tasks more accurately and recover from failure more reliably are the ones that reach usable success rates in production.

For enterprise buyers, a no-model-swap result lowers the barrier to improvement: instead of waiting for the next generation of models, they can invest in the agent architecture built on top of the model they already have.

What to watch next is how Salesforce applies this inside its own agent products, and whether the same approach transfers beyond the browser to desktop apps, internal systems and cross-system workflows. The report is a research result; whether it holds up in live customer environments will need more public evidence.