Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity. A key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. We therefore introduce The Trinity, a recursive self-improvement framework coupling world generation with Harness evolution through three roles: Creator, Explorer, and Evolver. Drawing on verified execution trajectories, Evolver shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. We introduce Capability-Driven Real2Sim2Real: for heterogeneous tasks at the same node, Creator abstracts their common requirements into simplified, verifiable simulated worlds for capability learning, while evaluation on original tasks measures transfer. Along this capability tree, Explorer progressively evolves reusable capability modules through selective inheritance, forming a Harness tree whose validated updates accumulate across iterations within a fixed topology. The resulting Harnesses specialize in different capabilities, motivating a Capability-Aware Harness Router that uses public task information to select compatible module configurations from the tree. Our evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
Figures & tables
Figure 2: Detailed pipeline of The Trinity. Three stages build task, capability, and Harness trees. Worlds are verified and screened; Harnesses evolve through selective inheritance and local refinement. Routing selects one compatible Harness and stores verified feedback for later tasks.
Figure 3: Capability-driven world creation at node d.
Table 1: Training and unseen benchmarks. * Agent-built from web-searched sources.
Figure 5: Benchmark-level Harness evolution. Left: Train; right: Unseen. Radial bars show Base and Iterations 1–4; * Agent-built from web-searched sources.
Figure 6: RSI mean performance across benchmarks. Horizontal lines show Codex and DeepSeek reference means.
Figure 7: Capability–protocol correlations across six protocols. Red is positive; blue is negative.
Harness
SOON (Sim)
CaP-X (Sim)
SOON (Real)
CaP-X (Real)
H0
0
10
0
0
H(1)
20
20
20
20
H(2)
20
30
20
20
H(3)
80
40
60
40
H(4)
80
60
60
60
Table 6: Embodied performance across iterations. Success rates (%) with prepared skills.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Task tree. The 44 benchmarks are grouped into five domains and 15 subclasses. Training benchmarks take the colour of their domain and unseen benchmarks a lighter shade of it; SOON and CaP-X show their simulation and real-robot versions; * marks environments rebuilt from published rules.
Benchmark
Measure, instances, episodes
Base H 0
H (1)
H (2)
H (3)
H (4)
Codex
DeepSeek
Train
AdaPlanBench [ 54 ]
accuracy, 3 queries, 27 episodes
22.2
70.4
74.1
74.1
74.1
–
–
Among Us [ 72 ]
win rate, 3 games, 27 episodes
11.1
48.1
63
63
63
–
–
ARC-AGI-3 [ 36 ]
success rate, 25 games, 183 levels
22.4
41
45.9
50.3
53.6
–
–
BoxingGym [ 44 ]
success rate, 3 environments, 45 episodes
60
80
84.4
84.4
84.4
–
–
Chess Tactics* [ 50 ]
mate found, 3 positions, 27 episodes
66.7
92.6
96.3
96.3
100
–
–
Appendix
Table 8: Full benchmark results. Scores on a 0–100 scale for the Base Harness H 0 , the Harnesses H (1) –H (4) after each RSI iteration, and the Codex and DeepSeek Harnesses on the unseen benchmarks; – marks runs that were not performed.
Figure 69: Embodied environments. Real scenes appear above their simulated counterparts, with SOON on the left and CaP-X on the right.
Figure 70: SOON with the Base Harness. Rows show simulation and real-robot views from the fixed office camera, followed by the head camera. The robot walks from the side room along the aisle, looks around, and turns towards the desk beside the fan, but reaches out before the desk is within reach in both settings.
Figure 71: SOON with the evolved Harness. Rows follow Figure 70 . In simulation the robot turns right past the doorway, walks diagonally towards the desk beside the fan, and slows to small steps before touching it. The real robot also approaches in small steps and completes the task.
Figure 72: CaP-X with the Base Harness. Rows show head, side, and wrist cameras in simulation, followed by the real robot. In both settings the robot manipulates other cups without inspecting the target cup and incorrectly reports that no cube is present.
Figure 73: CaP-X with the evolved Harness. Rows follow Figure 72 . In simulation the robot checks the cups and retries a missed grasp until the cube is visible. On the real robot it sweeps the cups aside to uncover the cube; both runs then report its colour and location.
Recent work extends recursive self-improvement (RSI) to agent harnesses for long-horizon coding and terminal tasks, enabling agents to improve execution mechanisms from experience. However, generalizable harness RSI remains challenging. First, evolving harnesses on evaluation benchmarks or their subsets makes it difficult to distinguish reusable improvements from benchmark-specific adaptation. Second, single-trajectory updates can conflate systematic harness deficiencies with instance-specific reasoning and solution details, producing modifications that transfer poorly to unseen tasks. Third, localizing recurring behavioral deficiencies within monolithic harnesses is difficult, while whole-harness optimization can entangle unrelated mechanisms and complicate attribution and validation. We propose ModularRSI, a benchmark-disjoint, contrastive, and modular framework for generalizable harness evolution. ModularRSI contrasts successful and failed trajectories for the same task and aggregates evidence across tasks to identify recurring behavioral deficiencies. It decomposes the evolvable harness into five functional modules: Agent Loop, Tool Use, Observation Management, Context Management, and Task Completion Detection. Each module evolves independently within a restricted modification scope, followed by an integration stage that combines the evolved modules into a unified harness and resolves potential conflicts. To support benchmark-disjoint evolution, we curate 2,000 executable evolution tasks from external sources that are disjoint from downstream evaluation benchmarks. Experiments on TB2.0 and SWE-Bench Verified show consistent improvements on unseen in-domain and cross-domain tasks, with the evolved harness also transferring across different foundation models.
Siwei Wu, Jincheng Ren, Yizhi Li +12
University of Manchester · IQuest Research · M-A-P +3
LLM agents' capabilities depend not only on their underlying models but also on their execution harnesses. Recursive self-improvement (RSI) offers a promising approach to automatically evolve these harnesses. However, deploying RSI in open-world settings still faces two critical challenges. First, open-world tasks are typically unseen during the harness design phase and novel at test time, leaving no ground-truth verifiers to reliably guide online harness edits. Second, the RSI process remains highly task-specific and fails to generalize to novel scenarios. We introduce Harness^2, a plug-and-play, recursive agent harnessing framework that addresses these challenges through two connected levels of recursion. Task-level recursion refines the current task's harness using a contrastive Propose-Probe-Compose step that reliably guides edits without ground-truth verifiers, and can be scaled further via parallel or sequential recursion. Harnessing-level recursion then extracts experience from these edits to update a continual improvement harness, enabling the agent to better adapt to future tasks. Across three professional benchmarks and four model-harness configurations, a single step consistently improves GLM-5 and Gemini 3.5 Flash over their base harnesses, yielding gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB, with further gains as the task-level recursion budget increases.
Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu +6
Northwestern University · Google Cloud AI Research
LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.
Minhua Lin, Juncheng Wu, Zijun Wang +14
1The Pennsylvania State University · 2UC Santa Cruz · 3Amazon +3