LAGO: Language Agents with Reusable Growing Options

1 Introduction

• A monolithic agent must decompose tasks, retrieve context, execute, verify, and recover within one trajectory, risking context overflow and accumulated errors.Software repair may require repository understanding, bug localization, patching, testing, and regression checks.Source: §1 · p. 1
• Static multi-agent pipelines can divide work across roles, but recurring sub-problems are still solved from scratch rather than stored, verified, and reused.Their roles, prompts, and interaction patterns are typically specified manually before execution.Source: §1 · p. 1
• LAGO addresses the challenge of choosing when to reuse an agent, spawn one, verify and admit it, and limit library redundancy.The evaluation also re-runs competing scaffolds on a shared 150-task subset under the same frozen policy and temperature.Source: §1 · p. 2
• LAGO treats specialized language agents as temporally extended options that a meta-agent can retrieve, spawn, evaluate, admit, and deduplicate.Each option includes a role specification, an LLM-tool policy, a termination condition, and a verifier.Source: §1 · p. 2
• The framework adds reusable agent behaviors to the meta-agent’s action space, alongside primitive tool calls.The library grows over a task stream through spawning, admission, reuse, and deduplication.Source: §1 · p. 2
• 58.6% of issues are resolved on the full 500-task split by LAGO with frozen Qwen3.5-35B-A3B in both roles.The evaluation uses single-attempt sampling at temperature 0 and the official Docker evaluator.Source: §1 · p. 2

2 Related Work

• ReAct interleaves reasoning and tool actions (Yao et al., 2023b), while tool-use systems often leave decomposition and verification inside one long trajectory.Related approaches include chain-of-thought (Wei et al., 2022), tree search (Yao et al., 2023a), self-reflection (Shinn et al., 2023), and direct or RL-based tool use (Schick et al., 2023; Singh et al., 2025; Jin et al., 2025; Wei et al., 2025b).Source: §2 · p. 2
• MetaGPT (Hong et al., 2024), AutoGen (Wu et al., 2023), and CAMEL (Li et al., 2023) coordinate fixed roles and prompts, while learned orchestration policies are also studied (Dang et al., 2025).LAGO instead stores reusable agent-options and lets its action space change across tasks.Source: §2 · pp. 2–3
• Figure 1 depicts decomposition, routing, verification, reuse, and deduplicated library growth as parts of LAGO’s reusable action space.Its panels illustrate a task, parallel option dispatch, and the option lifecycle.Source: Figure 1 · p. 3
• Self-evolving systems build code skills (Voyager; Wang et al., 2024), prompt-injected skills (CODESKILL; Li et al., 2026), repository memory (Wang et al., 2025), or libraries and episode-derived rules (Zhang et al., 2026; Tablan et al., 2026).LAGO reuses complete specialized agents, each with its own specification, termination condition, and verifier.Source: §2 · p. 3
• The options framework models temporally extended closed-loop policies, with later work learning or organizing such policies (Bacon et al., 2017; Vezhnevets et al., 2017).LAGO adapts the abstraction from motor control to language agents that make multiple tool calls before returning control to a meta-agent.Source: §2 · p. 3

3.1 Overview

• The meta-agent decomposes tasks into sub-tasks, assigns existing or newly spawned agents, and aggregates their results into a final answer or patch.Executor agents handle assigned sub-tasks through their own tool-use loops.Source: §3.1 · p. 3

3.2 Agents as Options

• Each agent-option packages a role specification, an LLM-tool policy, a termination condition, and a verifier for assessing the result.Termination occurs when the sub-task is completed, verified, failed, or reaches its execution limit.Source: §3.2 · p. 4

3.3 Decision State and Action Space

• The meta-agent’s decision state combines task state, active context, persistent memory, and the current option library; its actions include primitive tools and retrieved options.When no suitable option exists, LAGO can fall back to ordinary tool use.Source: §3.3 · p. 4

3.4 Spawn, Admission, and Governance

• LAGO spawns agents when the library lacks a suitable option, reuse repeatedly fails, or a task appears to need a new skill, then evaluates candidates before admission.Governance deduplicates similar agents, prunes rarely used or regression-failing agents, and versions each option’s prompt, tools, model, and verifier.Source: §3.4 · p. 4

3.5 Objective over a Task Stream

• Across task streams, LAGO’s objective balances task success and quality against library complexity, discouraging unbounded growth and unnecessary spawning.The library-size penalty encourages reuse over spawning.Source: §3.5 · p. 5

3.6 Improving the Meta-Agent (Optional)

• The optional supplementary analysis fine-tunes the meta-agent on successful decomposition traces while holding the executor fixed.The main system instead uses a frozen, untrained meta-agent.Source: §3.6 · p. 5

4.1 Experimental Setup

• The evaluation uses identical task IDs and a shared frozen policy, temperature, attempt count, and evaluator for its 150-task controlled scaffold comparison.The full 500-task result is not a row in that comparison, and approximate uncertainty at n=150 is ±8 percentage points; the cross-domain benchmarks test ordering rather than compare directly with the main result.Source: §4.1 · p. 5, Table 1 · p. 6

4.3 Component Ablation: Decomposition and Library Effects

• On a fixed 50-task split, disabling decomposition resolves 26 tasks (52.0%), shuffling reused content 28 (56.0%), and disabling the library 24 (48.0%).The full system resolves 30 tasks (60.0%); the authors interpret the effects as bounded evidence for a library role, not proof that decomposition alone is necessary.Source: §4.3 · p. 7

4.4 RQ2: Observed Differences Are Consistent with Decomposition Quality

• On SWE-bench Pro, LAGO reaches 50% versus 34% for the single-agent baseline; with DeepSeek-V3.2 fixed as executor, the SFT meta-agent reaches 40% versus 34% for the base variant.The authors describe the latter difference as consistent with a role for decomposition quality, while noting that prompts and control flow also differ.Source: §4.4 · p. 7
• On BrowseComp, dual-DeepSeek-V3.2 LAGO raises average accuracy from 33% to 46%, with English increasing from 22 to 32 and Chinese from 44 to 60.Swapping in Qwen3-30B meta-agent variants lowers accuracy; the result does not establish a domain-independent mechanism.Source: §4.4 · pp. 7–8
• The cross-model instance comparison finds the SFT meta-agent solving 20 SWE-bench Pro tasks to the base variant’s 17, with five new successes and two regressions.Wins cluster around base-agent under-decomposition, while regressions arise from over-specified scope that can cause missed edge cases.Source: §4.4 · p. 8, Table 6 · p. 8

4.5 RQ3: The Agent Library Is Actively Used and Deduplicated

• Across the three systems, 78.8%–86.4% of executor options are retrieved from the library rather than freshly spawned.In a separate 150-task rollout, the seeded library starts at 123 options, drops to about 80 after deduplication, and ends at 86; this does not establish long-run equilibrium across task distributions.Source: §4.5 · p. 8

5 Discussion

• Poor decompositions can create conflicting edits, miss verification, or set scopes that are too broad or too narrow; finer decomposition is not always better.SFT produces smaller, more verifiable sub-tasks with clearer dependencies, but a useful meta-agent must balance precision and coverage.Source: §5 · p. 8
• 84.4% of options were reused in the separate 150-task rollout (276 reuses / 51 spawns).The rollout began with the 123-option seed350 library, distinct from the 50-task reuse stream.Source: Figure 2 · p. 9
• The library fell from 123 options to about 80 after initial deduplication and ended at 86.This is the library-size trajectory for the same rollout as the reuse-rate figure.Source: Figure 3 · p. 9
• The marginal value of decomposition is task- and model-dependent; disabling it removed four resolved tasks on the 50-task ablation split.The cross-model results identify decomposition quality as a plausible factor in some complex agent tasks.Source: §5 · pp. 8–9
• Reuse alone is insufficient: the meta-agent must select an option that fits the sub-task, or a weak router can still fail.The library is actively reused, and deduplication limits unchecked post-initialization growth.Source: §5 · p. 9
• Under a fixed DeepSeek executor, the Qwen3 SFT meta-agent improves over the Qwen3 base, consistent with a contribution from decomposition quality.The broader cross-model ordering has dual DeepSeek strongest overall.Source: §5 · p. 9

6 Limitations

• At n=150, a resolve-rate difference carries roughly ±8 percentage points of binomial uncertainty, so only sizable Table 1 gaps are meaningful.The competing scaffolds were not rerun for the full-500 absolute result, so it cannot be merged with the 150-task comparison into one ranking.Source: §6 · p. 9
• Reuse is measured but not fully isolated, and the 50-task shuffle and library ablations are noisy and do not isolate transferred skill from generic prompt scaffolding.Spawning, deduplication, routing, and parallel versus sequential execution are also not individually isolated.Source: §6 · p. 9
• Library initialization must use tasks disjoint from evaluation to avoid contamination; verifier design is task-dependent and broader deployment needs more general verification.Meta-agent fine-tuning is optional, with moderate gains, and more data is needed to characterize when it helps versus over-specifies.Source: §6 · p. 10

7 Conclusion

• LAGO reuses and spawns specialized agent-options, verifies outcomes, and maintains a growing library without weight updates.On SWE-bench Verified, frozen Qwen3.5-35B-A3B in both roles resolves 58.6% of the full 500 tasks; the controlled comparison reruns competing scaffolds on the same policy and 150-task subset.Source: §7 · p. 10
• The results suggest decomposition is most useful when sub-problems are separable and recur across the task stream; transfer to other growing-library systems remains empirical.Source: §7 · p. 10

AI use statement

• The authors report using generative AI for code editing, figure scripting, language editing, and manuscript review, with outputs reviewed and code and figures checked against underlying data.They state that AI tools were not used to generate experimental results or citations.Source: AI use statement · p. 10

A Inference Loop

• The loop starts from an initial option library and decomposes each task into sub-tasks.Source: A Inference Loop · p. 13
• For each sub-task, it retrieves candidate options and reuses a suitable one or spawns an option when none fits.Source: A Inference Loop · p. 13
• Selected options are executed and verified; successful novel options are admitted, and verified sub-task results are aggregated into an answer or patch.Source: A Inference Loop · p. 13
• The library is periodically deduplicated, pruned, and versioned as the task stream proceeds.The procedure is placed here to keep the main text focused on the mechanism and its evidence.Source: A Inference Loop · pp. 12–13

• The appendix houses supporting figures and tables referenced from Section 4.4 to keep the main-text evidence focused on the frozen Qwen3.5-35B-A3B comparison.Source: B Additional Experimental Results · p. 13

• Methods using the same Qwen3.5-35B-A3B policy report on different task selections and grading setups, so their numbers are listed for provenance rather than ranked against the re-runs.Hand-selected rather than stratified subsets and CODESKILL’s partial-credit soft evaluator can raise reported numbers relative to the binary task-id-matched metric.Source: C Reference Methods Under… · p. 13
• 59.8–61.6% native-context results for mini-SWE-agent and Terminus 2 are retained for provenance, but their gap from re-run results does not isolate context length.Deployment, context, and inference conditions differ.Source: C Reference Methods Under… · p. 13
• 50% resolve rate for dual DeepSeek-V3.2 on SWE-bench Pro is a 16-point gain over the single-agent baseline.With the executor fixed to DeepSeek-V3.2, the SFT Qwen3-30B meta-agent improves 6 points over the base variant, which matches the baseline.Source: Figure 4 · p. 14

• Fine-tuning uses success- and decomposition-quality-filtered traces with task, decomposition, sub-task specifications, and verification-oriented instructions.The Qwen3-30B meta-agent is trained with supervised fine-tuning using 4-bit QLoRA.Source: D Meta-Agent Fine-Tuning Details · p. 14
• Fine-tuning targets step completeness, clearer sub-task chains with explicit dependencies, and precise task naming, but overly narrow sub-task scope can miss edge cases.Source: D Meta-Agent Fine-Tuning Details · p. 14
• 17 to 20 solved instances is the net gain for the SFT Qwen3-30B meta-agent over the base variant on the 50-task SWE-bench Pro split.The systems jointly solve 15 and jointly fail 28 instances; SFT uniquely solves 5 and regresses on 2.Source: Figure 5 · p. 15
• Figure 6 tracks pre- and post-deduplication library size, agents merged, and net growth per batch for a separate 50-task Dual DeepSeek stream.This stream is distinct from the 150-task Qwen3.5-35B-A3B rollout in the main text.Source: Figure 6 · p. 15
• On the 150-task subset, most tasks reach a submit, 31 more are recovered by the unconditional final-diff fallback, and 27 end with no patch.Among no-patch cases, dominant causes include malformed-tool-call retries, abandonment after a sub-agent reaches its execution limit, and turn-limit exhaustion.Source: Figure 7 · p. 16
• Table 7 lists subset-evaluated training-free methods using the same Qwen3.5-35B-A3B policy for provenance, while RL-trained methods are outside the weight-frozen scope.TDAD’s 25-task stage uses a 4-bit quantized policy; its 100-task stage uses Qwen3-Coder-30B, not this model.Source: Table 7 · p. 16