A new arXiv paper from researchers at the University of Hong Kong and Salesforce AI Research trains an AI agent and its harness at the same time, with reinforcement learning, and reports 97.0% success on the ALFWorld benchmark. The method, called JoMoHa, averaged 70.3% across ALFWorld, WebShop and DBBench, the highest of the methods the authors evaluated.
The harness is the code around a model that decides how its abilities are used. In JoMoHa one policy writes that harness as an executable program and a second policy solves the task inside it; both learn from the same task outcome. Both are Qwen3.5-9B-class models. By our arithmetic from the paper’s table, the 70.3% average is 22.4 points above the same task model using GPT-5.5 as a fixed harness editor.
Business Pill · AI Agents Explained
An evergreen Business Engineer explainer on AI agents and the work they automate. It teaches the general idea only and does not discuss this paper.
The key insight: A paper from the University of Hong Kong and Salesforce AI Research trains the harness-writing model and the task-solving model together with reinforcement learning and reports 97.0% on ALFWorld and a 70.3% three-benchmark average; by our arithmetic that is 22.4 points above GPT-5.5 used as a fixed harness editor.
The paper calls the two policies the Engineer and the Solver. The Engineer generates an executable harness; the Solver interacts with the environment under it. Each is trained on its own generated tokens from the shared task result. The authors summarise it as: “One model learns to write the runtime; another learns to act under it.”
The paper says WHALE, an earlier method, still relies on a fixed external proprietary frontier model to propose harness improvements, while Harness-R1 trains the harness writer but keeps the task model frozen. Elvis Saravia, who flagged the paper on X, put it this way: the harness editor is usually a fixed frontier model. JoMoHa trains that editor too. The code is released on GitHub.
Average success on ALFWorld, WebShop and DBBench by method, with Qwen3.5-9B as the task model in every row. Data: JoMoHa paper, arXiv 2611.11499, Table 4; baselines other than JoMoHa and WHALE are taken by the authors from Harness-R1. The 22.4-point gap is our arithmetic.
What Was Posted on X
Nice read from Salesforce AI Research and colleagues.
There is a lot of interest in self-improving harnesses, but the harness editor is usually a fixed frontier model.
JoMoHa trains two policies with RL from the same task outcomes. One writes the executable harness, and the other solves the task inside it.
Updating the model and the harness in turns caused trouble. Performance rose and then fell as the two became incompatible. Keeping separate histories and value estimates for each policy fixed that.
JoMoHa beat weight-only and harness-only training on ALFWorld, WebShop, and DBBench, averaging 70.3% and reaching 97.0% on ALFWorld. Training the harness writer also beat freezing it.
Elvis Saravia posted this at 00:20 UTC on 13 November, summarising the JoMoHa paper and linking it on arXiv. The post sets out the idea (one policy writes the harness, the other solves the task inside it), the instability of taking turns and the headline results. This piece takes every figure from the paper itself.
Why Taking Turns Broke
The authors first tried the obvious approach: alternate updates to the two policies. On ALFWorld that run rose from 83.0% to 90.5%, then fell across later stages to 48.5% and 31.5%. By our arithmetic that is a 59.0-point drop from its best stage. The paper describes this as progressive incompatibility: each side adapts to a partner that then changes.
Holding one saved harness fixed, the authors split the damage. With the later Solver, the same harness cost 17.0 points instead of 4.0, and the Solver alone also lost 34.5 points without any harness. In one case the later Solver finished a task in 14 turns without the harness but hit the 50-turn limit with it. The cause was a lookup bug in the generated code: “A dotted string is a different dictionary key, not a nested lookup, so the completed action is missed.”
JoMoHa’s fix keeps the two learners coordinated: separate value estimates aware of the partner, role-specific return normalisation, partner-trajectory weighting and asymmetric update schedules. Each of these alone lifted a joint-only baseline from 88.5% to between 91.5% and 92.5%; all four together reached 97.0%, the paper reports.
The Benchmark Table
JoMoHa scored 97.0% on ALFWorld, 45.8% on WebShop and 68.0% on DBBench. The paper says its average exceeds Agent SFT + Harness-R1 by 6.1 percentage points and WHALE by 7.2 points. Six fixed frontier models used as harness editors, including GPT-5.5, GLM-5.2 and DeepSeek-V4-Pro, averaged between 44.3% and 48.8% with the same 9B task model; the best of them, GLM-5.2, sits 21.5 points below JoMoHa by our arithmetic.
Training the harness writer mattered. A Frozen Engineer control, which kept sampling fresh harnesses from the initial model, reached 90.5% on ALFWorld; weight-only training reached 89.5% and harness-only 80.5%. The largest gains were on Clean, Heat and Cool tasks, which require tracking whether an object’s state has changed.
As we read it, the paper treats the harness as something learned alongside the model rather than written once around it. Its own diagnosis shows why that matters: when the two took turns, a harness that worked for one Solver checkpoint cost the next one 17.0 points, and the whole run fell to 31.5%. The result says compatibility between model and harness has to be maintained, not assumed.
Elvis Saravia on X, 13 November 2026
“There is a lot of interest in self-improving harnesses, but the harness editor is usually a fixed frontier model.”
Three Implications
FOR AGENT BUILDERS The paper reports that a harness saved during training raised a fixed Solver from 92.5% to 96.5% on ALFWorld without changing model weights.
FOR MODEL LABS Six frontier models used as fixed harness editors averaged between 44.3% and 48.8% with the same 9B task model, against 70.3% for JoMoHa.
FOR TEAMS UPDATING MODELS In the alternating run the same harness cost a later Solver 17.0 points versus 4.0 for an earlier one, the paper says.
The analysis says: “the conversational interface was scaffolding. The harness is the building.”
As we read it, JoMoHa pushes that idea into training itself: the paper makes the harness a policy that learns from task outcomes. The analysis predates the paper and does not discuss it.
ALFWorld success by stage of naive alternating training, each stage pairing the Solver with its own harness, versus JoMoHa after 144 rounds. Data: JoMoHa paper, arXiv 2611.11499, Table 2; the 59.0-point drop is our arithmetic.
What Is Not Established
This is a preprint, not peer reviewed. Except for JoMoHa and WHALE, which the authors ran themselves, baseline scores are taken from the Harness-R1 paper, using its training and test sets. The detailed coordination and ablation results are on ALFWorld only. The results cover 9B-class task models; how the method behaves with larger task models, and at what training cost, is not established.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
JoMoHa trains a harness-writing policy and a task-solving policy together from the same reward and reports 97.0% on ALFWorld and a 70.3% three-benchmark average. Naive turn-taking collapsed to 31.5%; keeping the learners coordinated is what the authors credit for the gain. By our arithmetic the average is 22.4 points above GPT-5.5 used as a fixed harness editor.
A note on sourcing. We read the JoMoHa paper (arXiv 2611.11499, by Xudong Wu, Guangchen Lan, Mingqi Yuan and Jiayu Chen of the University of Hong Kong and Salesforce AI Research) in full, including its benchmark and ablation tables, and Elvis Saravia’s post on X, which was passed to us. All results are the authors’ own; we did not rerun them. The 22.4-point, 21.5-point and 59.0-point gaps are our arithmetic from the paper’s tables. We did not contact the authors, the University of Hong Kong or Salesforce.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
