Quick Facts
- Salesforce’s DarwinX framework lifted an AI agent’s browser task completion from 43.5% to 93% without modifying model weights.
- Invalid agent trajectories dropped from 23.5% to 1.4% in the same experiments, run on 1,260 unseen tasks.
- The open-source Beagle framework, which ships DarwinX as its core algorithm, was released on Sept. 2, 2026, under the Apache 2.0 license.
Salesforce AI Research has published a framework that more than doubles AI agent performance on browser tasks by changing only the scaffolding around a language model, leaving the model itself untouched. The system, called DarwinX, treats agent optimization as a selection problem borrowed from evolutionary biology.
In tests on WebArena-Infinity, the team evolved a harness on 300 synthetic tasks drawn from application documentation, then evaluated the result on 1,260 unseen tasks with deterministic verifiers. With GPT-5.5 frozen, the agent’s verified completion rate climbed from 43.5% to 93%. Cases where the agent produced unusable outputs fell from 23.5% to 1.4%.
What DarwinX Changes
A language model agent depends on more than its weights. It also depends on its harness: the prompts it reads, the tools it can call, the notes it keeps, and the control flow that sequences its actions. DarwinX proposes small edits to that harness, tests each edited agent on real tasks, and keeps only changes that solve something new without breaking what already worked.
The system maintains a population of harness variants and applies natural selection across them. Fitness scores come from each benchmark’s own verifier. No gold solutions are hand-picked. An archive of previous lineages supplies material for recombination, preventing the system from trading old capabilities for new ones.
Senior author Ran Xu identified the problem with conventional manual tuning. “Manual harness engineering can easily converge to a local optimum: you fix the prompt or workflow for one failure mode, but without broad regression testing, you may silently break something that was already working,” Xu said.
Results Across Four Benchmarks
DarwinX improved scores on all four benchmarks the team tested. Gains ranged from 3.4 points on SWE-bench Verified to 49.5 points on WebArena-Infinity. On average, one evolution loop added about 17 points.
The team also tested transfer. They took a harness evolved on Terminal-Bench 2.1 and ran it unchanged on SWE-bench Verified. It reached 84.2%, compared with 80.8% for the reference harness, despite receiving no SWE-bench feedback during evolution. The result suggests that some harness improvements carry over to benchmarks they were never trained against.
An audit of every WebArena-Infinity trajectory confirmed the gains held under scrutiny. The base agent’s raw score of 53.0% fell to 43.5% on closer review. The DarwinX agent’s 94.4% held at 93.0%. The gap between the two widened rather than narrowed when audited.
Business Implications for Enterprise Builders
DarwinX is directly relevant to software teams that build on hosted models and have no fine-tuning pipeline. Because the method never touches model weights, any team calling a commercial API can apply it. The approach does not require access to model internals.
Xu described the broader goal for enterprise teams: “Enterprises do not necessarily need a single perfect reward function. They need an infrastructure layer that continuously turns real work into increasingly strong enterprise intelligence.”
Production deployments rarely arrive with reliable verifiers attached to every task. The Salesforce team addressed this by recommending teams maintain a proxy evaluation suite built from the workflows their agent actually encounters, then run evolution against that suite periodically and deploy only versions that clear regression checks.
Open-Source Release
The paper, authored by a 12-person team led by Yifan Zhang, Yutong Dai, and Juntao Tan, was published on arXiv on July 31, 2026. Salesforce released Beagle, the open-source framework that ships DarwinX as its core algorithm, on Sept. 2, 2026. It is available on GitHub under the Apache 2.0 license. Teams can bring their own agent harness and verifier-backed tasks into Beagle. Salesforce’s proprietary Monet agent remains closed.