acceptodds
Under review as a conference paper at ICLR 2027

LAGO: Language Agents with Growing Reusable Options

Abstract

Language agents based on monolithic tool-use loops struggle with long-horizon tasks that require decomposition, specialized execution, and verification, while conventional multi-agent systems rely on fixed, manually designed roles. We introduce LAGO, a meta-agent framework that represents specialized language agents as reusable temporally extended options. The meta-agent retrieves suitable options from a persistent library or spawns new ones, dispatches them to solve sub-tasks, verifies their outcomes, and admits successful and novel options for future reuse. This turns task decomposition into a reusable capability and allows the effective action space to grow across a task stream without updating model weights. On the full 500-task SWE-bench Verified benchmark, LAGO using frozen Qwen3.5-35B-A3B in both the meta-agent and executor roles achieves a 58.6% resolve rate under single-attempt evaluation with the official Docker evaluator. Cross-model and cross-domain experiments further support the role of decomposition quality, while library analysis shows frequent option reuse and bounded growth under deduplication. These results suggest that reusable agent-options are a promising abstraction for language agents that accumulate reusable task-solving capabilities over time.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Language Agents, Multi-Agent Systems, Task Decomposition, Tool Use

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_cJGtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces LAGO (Language Agents with Reusable Growing Options), a framework that treats specialized language agents as reusable, temporally extended actions ("options"). For each task, a meta-agent: breaks the task into sub-tasks; for each sub-task, either retrieves a suitable agent from a persistent library or spawns a new one; verifies each sub-agent's result; adds newly spawned agents to the library if they succeed and are novel. The library is maintained through deduplication, pruning and versioning, and model weights are never updated. With frozen Qwen3.5-35B-A3B as both meta-agent and executor, LAGO resolves 58.6% of SWE-bench Verified. On a 150-task subset it scores 60.3%, compared with 57.3% for Terminus 2 and 49.3% for mini-SWE-agent. The paper also reports results on SWE-bench Pro and BrowseComp, and an optional fine-tuning (SFT) step for the meta-agent.

Critical strengths
  1. Treating specialized agents as reusable actions, in a library that grows over a stream of tasks, is a sensible way to avoid solving recurring sub-problems from scratch.
  2. Disabling decomposition or the library lowers performance, which offers some evidence that both components help.
Critical weaknesses
  1. The agent library is presented as the main contribution, yet its core operations are not defined. The admission rule on Line 184 depends on a "novelty" score​. The novelty metric is never defined, and the threshold values are never given. For deduplication and pruning, the paper does not say how "similar specifications and behavior" is measured, how two agents are merged, what counts as "rarely used," what the regression checks are, or how often these steps run. For routing, it says only that retrieval uses "a similarity function over option specifications," without giving the similarity measure, the number of candidates, or the threshold for a "suitable" option. The reuse-or-spawn decision appears to be left to the LLM's judgement. Finally, all experiments start from a 123-option seed library, but the paper never explains how it was created. It says only that it came from tasks disjoint from the evaluation set.
  2. The paper is hard to follow, and the figures and captions need significant polishing. There are formatting errors (e.g., Line 239). Internal identifiers appear without explanation, such as "max_branch_turns" in Figure 2 and "agent_library_seed350" in Lines 393–395. Important details are also spread across the main text and the appendices.
Questions for authors
  1. How exactly are novelty, the similarity used for retrieval and deduplication, and the admission and pruning thresholds computed? What values were used?
  2. How was the seed library built? How does performance change when starting from an empty library?
  3. What is the inference cost (tokens and LLM calls) of LAGO compared with the baselines?
Reviewer_tGgROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

LAGO is a design method for a two-tier LLM agent system: the top tier is a meta-agent, and the lower tier consists of multiple agents. The meta-agent maintains a library of agents by storing, adding, and spawning agents. Whenever an agent is needed, the meta-agent first selects an appropriate one from the library; if no suitable agent exists, a new agent is created, and it is added to the library if it is verified to be capable of the task. The agents in the library are also updated regularly (deduplication, pruning, and versioning).

Experiments:

  • SWE-bench Verified: both the meta-agent and the agents are driven by Qwen3.5-35B-A3B.
  • SWE-bench Pro and BrowseComp: the models used are DeepSeek-V3.2, Qwen3-30B, and a fine-tuned Qwen3-30B.
Critical strengths

The method is sound, intuitive, simple, and concise.

The idea is intuitively promising: a good library of well-defined, task-specific agents is highly likely to improve the performance of an agent system.

Critical weaknesses

The originality may be insufficient

The method largely overlaps with existing work, for example:

  • Meta-agent + agents: MetaGPT, AutoGen
  • Verification before admission, and retrieval based on description similarity: Voyager
  • Self-expanding agent library: ADAS (ICLR 2025), Darwin Gödel Machine (ICLR 2026)
  • Accumulating experience on "training" tasks and reusing it on "test" tasks: ExpeL
  • Agentic self-improvement: Self-Harness, Meta-Harness

LAGO overlaps substantially with the work above, so its novelty is limited. Moreover, the authors cite only part of this work (ADAS, Darwin Gödel Machine, ExpeL, and Meta-Harness are not cited; Self-Harness appears only in Table 7 of the appendix).

As I understand it, the main contribution and novelty of this work is the method for managing and maintaining the agent library. However, many commercial and open-source products already support agent templates and sub-agents, for example Claude Code (https://code.claude.com/docs/en/sub-agents) and OpenClaw (https://docs.openclaw.ai/tools/subagents). One can simply add a well-designed prompt to the main memory file (e.g., CLAUDE.md, AGENTS.md, MEMORY.md), and these tools already work well at managing agents. If so, why should LAGO's harness-level design be preferred over a simpler prompt-level approach? I believe the authors should discuss this.

The main experiment may not be able to support the claims

The inference settings of the main experiment are:

  • Temperature: 0
  • Context length: 65,536 tokens (Sec. 4.3; the baselines are re-run under LAGO's context budget, Sec. 4.2)

However, the official Qwen3.5-35B-A3B model card recommends a temperature of 0.6–1.0 in thinking mode and 0.7 in non-thinking mode, and a context window of at least 128K. Moreover, a single-agent baseline must keep its entire context in one window, whereas in LAGO every agent has its own independent context window.

These settings coincide with much lower scores for the baselines. In Sec. 4.2 and the appendix, the paper cites previously reported results on the same backbone (from HarnessBridge): 59.8 for mini-SWE-agent and 61.6 for Terminus 2. Under the above settings, however, the experiments yield only 49.3 and 57.3.

In addition, the authors do not report other inference parameters, such as top-p or whether thinking mode is enabled.

Overall, the experimental comparisons are not fair.


These weaknesses, especially the first one, raise my major concerns, so my current inclination is to reject. I look forward to hearing back from the authors; as long as these issues can be resolved, I would be very glad to raise my score.

Questions for authors
  1. Could you explain the originality of your method in more detail, especially what makes it unique and more effective or efficient than existing methods?
  2. What is your resource usage compared with the baselines? For example, keeping many sub-agent processes alive requires extra memory (host RAM, not GPU memory); if sub-agents are discarded once their jobs are done, every respawn requires a new prefill. What are these costs, and how do you balance them?
  3. How is the pre-defined agent library designed, how is each agent obtained, and how much computation does it cost to build this initial library?
  4. Could you re-run the experiments with the officially recommended parameters and a fair context window length?

If these questions are well answered and resolved, I believe the paper will become solid in both correctness and originality.

Reviewer_bFBvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LAGO, a two-level agent framework that treats specialized language agents as options. A meta-agent decomposes a task, retrieves agent-options (role specification, tool loop, termination condition, verifier) from a persistent library or spawns new ones, and admits verified, novel options; the main system updates no weights. With a frozen Qwen3.5-35B-A3B in both roles and one attempt, LAGO resolves 58.6% of SWE-bench Verified. On a fixed 150-task subset, re-runs of mini-SWE-agent, Terminus 2 and LAGO on the same policy give 49.3%, 57.3% and 60.3%. Cross-model runs on SWE-bench Pro and BrowseComp vary the meta-agent and executor, with dual-DeepSeek LAGO beating a single-agent baseline (50 vs 34; 46 vs 33), and a separate rollout reports 84.4% option reuse.

Critical strengths

The problem is real: a single ReAct loop has to localize, edit, test and recover inside one context, and recurring sub-problems get solved from scratch on every issue. Storing verified sub-agents and reusing them across a task stream is a reasonable thing to try.

Critical weaknesses

The biggest concern is that the reusable growing options the paper is named after do not seem to be doing the work, and I see little reason to expect they would on SWE-bench.

In my view what recurs across Verified issues is generic procedure (localize, edit, run tests) that a decent scaffold prompt already encodes. A stored option is a prompt-level specification run by the same frozen model, not executable code as in Voyager [1], and the admission thresholds θ_task, θ_nov are never given.

The paper's own numbers fit this reading: shuffling reused content costs 2 of 50 tasks, the measured library drops from 123 seeded options to about 80 and ends at 86 despite 51 spawns, and nothing plots resolve rate against stream position, the test of "more capable on future tasks".

Experiment-wise, each LAGO sub-agent runs its own tool loop, which the paper credits with reducing context pressure, and no token or wall-clock costs are reported. Which library do the Table 1 and full-500 runs start from? The seed is "disjoint from the evaluated subset", but the rollout uses agent_library_seed350 and 500 − 150 = 350, so that seed may overlap the full-500 run.

[1]. Voyager: An Open-Ended Embodied Agent with Large Language Models

The RQ1 headline rests on a 3-point gap over Terminus 2, which by the paper's own ±8-point variance is not meaningful.

Questions for authors

Please see above, the empirical evidence seems weak.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.