LAGO: Language Agents with Reusable Growing Options

Yuran Wu; Mengkang Hu; Qionglin Qiu; Fanqing Meng; Qiguang Chen; Yuhang Zhou; Qiudan Zhang; Xu Wang

cs.AIOct 14, 2026

LAGO: Language Agents with Reusable Growing Options

Authors: Yuran Wu, Mengkang Hu, Qionglin Qiu, Fanqing Meng, Qiguang Chen, Yuhang Zhou, Qiudan Zhang, Xu Wang

Organizations: Shenzhen University · University of Hong Kong · Hunan University · National University of Singapore ·

Abstract

Long-horizon tasks expose a weakness of monolithic tool-use loops: one agent must decompose, execute, and verify a task within a single trajectory. We introduce LAGO, which treats specialized language agents as reusable temporally extended options. A meta-agent retrieves options from a persistent library or spawns new ones, dispatches sub-tasks, verifies results, and admits successful novel options. The library grows across tasks without updating model weights. On 500 SWE-bench Verified tasks, frozen Qwen3.5-35B-A3B in both roles resolves 58.6% under single-attempt evaluation with the official Docker evaluator. On a fixed 150-task subset, we re-run mini-SWE-agent, Terminus 2, and LAGO with the same policy, temperature, task IDs, and evaluator; their resolve rates are 49.3%, 57.3%, and 60.3%, respectively. Cross-model experiments provide evidence consistent with a role for decomposition quality. A 50-task component ablation finds four fewer resolved tasks without decomposition, two fewer with shuffled reused content, and six fewer without the library. These results support reusable agent-options as a practical abstraction for task streams, while leaving component-level causality task- and model-dependent.

Figures & tables

Figure 1: LAGO turns task decomposition into a reusable action space. (a) An illustrative SWE-bench-style issue is decomposed, routed to reusable or newly spawned options, executed, verified, and reused on a later task. (b) The overall framework shows the meta-agent dispatching parallel options and receiving verified results. (c) The option lifecycle governs specification, execution, verification, admission, and deduplicated library growth from $L_{0}$ to $L_{T}$ .

Frozen Qwen3.5-35B-A3B; 150-task subset; temperature $0$ ; single attempt; official Docker
Method	Mechanism	Resolve (%) $\uparrow$	$\Delta$ vs. mini-SWE $\uparrow$
mini-SWE-agent	single-agent ReAct loop	49.3 R	–
Terminus 2	terminal-agent scaffold	57.3 R	+8.0
LAGO (ours)	hierarchical meta-agent with growing options	60.3 R	+11.0

Table 1: Controlled scaffold comparison on a 150-task subset of SWE-bench Verified (repo-stratified, seed 42; identical task ids in every row). All rows use the same frozen Qwen3.5-35B-A3B at temperature

0

, one attempt per task, and the official Docker evaluator.

\Delta

is the descriptive percentage-point difference from mini-SWE-agent in this run. Each row is one end-to-end rerun ( R ) under the shared task protocol; Terminus 2 is run at temperature

0

for comparability.

Configuration	Resolved	Resolve (%)	$\Delta$ vs. full
LAGO (full)	30/50	$60.0\pm 13.6$	–
Disable decomposition	26/50	$52.0\pm 13.8$	$-4$
Shuffle reused content	28/50	$56.0\pm 13.8$	$-2$
Disable library	24/50	$48.0\pm 13.8$	$-6$

Table 2: Component ablation on a fixed 50-task split. All rows use denominator 50; unfinished rollouts are unresolved in this end-to-end accounting. The

\pm

values are approximate 95% binomial intervals, and

\Delta

is the absolute difference in resolved tasks.

Table 4

Table 5

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: SWE-bench Pro resolve rate by configuration (visualizing the numbers in Table 4 ). LAGO with DeepSeek-V3.2 as both meta-agent and executor reaches 50%, a 16-point gain over the single-agent baseline. With the executor fixed to DeepSeek-V3.2, an SFT Qwen3-30B meta-agent improves over the base variant by 6 points, while the base variant matches the baseline.

Figure 5: Instance-level comparison of base and SFT Qwen3-30B meta-agents on the SWE-bench Pro 50-task split (DeepSeek executor). The two systems share 15 jointly-solved and 28 jointly-failed instances; SFT uniquely solves 5 and regresses on 2, for a net gain of 3 instances (17 to 20 solved).

Figure 6: Agent library dynamics for a separate 50-task Dual DeepSeek stream, distinct from the 150-task Qwen3.5-35B-A3B rollout in the main text. Left: library size before and after deduplication (red labels show agents merged). Right: agents merged and net growth per batch.

Figure 7: Failure modes on the 150-task subset. (a) Patch provenance: most tasks reach a submit , a further 31 are recovered by the unconditional final-diff fallback, and 27 end with no patch. (b) Among those 27, the dominant causes are malformed-tool-call retries, abandonment after a sub-agent reaches its execution limit, and turn-limit exhaustion.

Method	Eval set	Orig. (%)	Setup note
Self-Harness	100 tasks	41.5	self-modifying harness; mean of 2 runs; 67 held-in / 33 held-out
CODESKILL ( Li et al., 2026 )	150 tasks	57.3–66.0	frozen policy $+$ trained 4B skill manager; soft evaluator $V(\tau)\in[0,1]$
CoACT	200 tasks	57.0–60.5	mini-SWE-agent $+$ observation compression
TDAD	25 tasks	24.0–32.0	4-bit quantized policy $+$ OpenCode; test-driven skill

Appendix

Table 7: Training-free (weight-frozen) methods on SWE-bench Verified that use the same Qwen3.5-35B-A3B policy but evaluate only on a subset , listed for provenance and not ranked against our re-run results (Self-Harness: arXiv:2606.09498, v3; CODESKILL: arXiv:2605.25430; CoACT: arXiv:2607.02911; TDAD: arXiv:2603.17973, whose 25-task stage uses a 4-bit quantized policy—its 100-task stage uses Qwen3-Coder-30B and is not this model). Ranges span each method’s reported variants. RL-trained methods (e.g., LEGO-RL, CompactionRL) are outside our weight-frozen scope and omitted.

Long-Horizon LLM Agents

Long-horizon large language model agents plan, use tools, and preserve task state across extended, multi-step interactions.

See all 451 papers

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks
Aug 3, 2026Ziyu Ma, Hailang Huang, Shun Zou +5Long-Horizon Agent Evaluation Long-Horizon LLM Agents
DreamX Team, Alibaba Group
DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents
Oct 14, 2026Tianle Chen, Pushkar Singh, Ketan Todi +8Long-Horizon LLM Agents Self-Training for Language Models
Google · Boston University · University of Washington +2
SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research
Jun 8, 2026Pu Ning, Quan Chen, Kun Tao +7Multi-Agent LLM Systems LLM Agent Orchestration
Tsinghua University · Peking University · Ant Group +1

Aug 3, 2026J/K move · Enter open · S save

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

Ziyu Ma, Hailang Huang, Shun Zou, Yong Wang+4

DreamX Team, Alibaba Group

Long-Horizon Agent Evaluation Long-Horizon LLM Agents LLM Agent Harnesses

Open paper AskPDF arXiv

Large language model (LLM) agents increasingly undertake long-horizon tasks that require sustained reasoning, tool use, and revision across many interdependent steps. However, existing agent harnesses maintain task execution, task state, and completion assessment within a growing context, making the state difficult to track and allowing incorrect self-assessments to propagate into later decisions. We reformulate long-horizon execution as a task-state management problem and propose LongHorizon-Harness, which maintains the task state explicitly outside execution and updates it only with facts independently verified from the environment. Its Manage-Execute-Audit(MEA) loop uses a manager to maintain the task state and determine the next subtask, a fresh-context executor to perform it, and a read-only auditor to verify the resulting environment state before the next round. A lightweight AgentAdapter supports interchangeable model and harness backends without modifying their native agent loops. LongHorizon-Harness improves Qwen~~3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench~~2.1, and from 2.8% to 8.3% on OSWorld~~2.0. It also raises Claude Opus~~4.7 from 20.0% to 34.3% on an OSWorld2.0 subset, demonstrating consistent gains across models, harnesses, and interaction domains.

Abstract

Long-Horizon LLM Agents

Explore similar work

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents

SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research