Language agents get a second trainable policy: JoMoHa learns the harness too, averaging 70.3% success across three environments

An agent's capability is no longer set by model weights alone: the policy that writes the runtime program becomes a trainable object and learns from the same task attempt as the policy that executes the task. Earlier routes either trained only the task-performing weights or delegated harness improvement to a fixed external frontier model, where execution feedback selected harnesses but never updated the proposer itself. On ALFWorld, WebShop and DBBench, JoMoHa reaches 97.0%, 45.8% and 68.0% success, an equally weighted average of 70.3%, 6.1 points above Agent SFT + Harness-R1 and 7.2 points above WHALE.

Both policies share the reward of one task attempt: the Engineer learns which harnesses support successful execution, and the Solver learns to complete tasks under them, each updated on its own generated tokens. The unit of learning is a complete task attempt, comprising the generated harness, the actions taken under it and the final outcome; an invalid proposal falls back to the base executor instead of failing the round. The ALFWorld coordination study uses separate Qwen3.5-9B-class Engineer and Solver actors with independent critics, initialized from Agent SFT, collecting eight paired episodes per round for 144 rounds.

The two roles differ in output length and decision horizon, so their returns differ in scale under a shared reward; critic targets are normalized per role, old experience is reweighted by the partner trajectory's likelihood ratio, and the Solver receives extra rounds that update only itself. A Joint-only version that trains both policies without these four coordination mechanisms ends at 88.5% success, while adding Context, Schedule, Peer or ValueNorm separately raises it to 91.5% to 92.5%. With all four mechanisms the success rate reaches 97.0%, 8.5 points above Joint-only; Context alone gives the highest single-addition point estimate at 92.5%.

Evaluation covers ALFWorld, WebShop and DBBench, testing household task completion, goal-directed shopping and database interaction respectively; the ALFWorld coordination study uses Qwen3.5-9B-class Engineer and Solver actors initialized from Agent SFT, with eight paired episodes per round over 144 rounds. Except for locally evaluated JoMoHa and WHALE, baseline scores are taken from Harness-R1 using its training and test sets, with checkpoint and harness selection on a separate validation set. At deployment, a saved harness is selected and evaluated with a frozen Solver checkpoint.

Alternating training on ALFWorld first rises from 83.0% to 90.5% and then falls stage by stage to 31.5%; fixed-component comparisons split the decline into lost Solver capability and a harness that becomes more harmful as the model changes, but the latter is described as consistent with mismatched harnesses steering the Solver toward less effective behavior, not as a confirmed cause. Pick2 remains the weakest ALFWorld category at 91.8%, leaving room when multiple objects must be collected and placed.