Long-horizon problem solving in artificial intelligence has traditionally been approached by scaling model parameters or extending execution horizons. Agents-A1, a 35B Mixture-of-Experts agent model introduced in prior work, improved long-horizon task performance by scaling the agent horizon rather than model parameters. It built a knowledge–action infrastructure that connected external knowledge, tool calls, observations, and verification feedback, producing training trajectories with an average length of 45K tokens. While Agents-A1 showed strong performance in multi-step search, scientific reasoning, instruction following, and tool use, subsequent research revealed that strong domain skills alone are insufficient for complicated tasks requiring subtask management, approach comparison, and decision revision as new evidence becomes available. A single agentic model working step by step may struggle with growing context, slow exploration, and errors that compound across decisions. Longer execution does not always lead to better results: an agent may repeat unproductive searches, build on unsupported assumptions, or accept a conclusion before checking all evidence. To address these limitations, Sage-A1.5, a 35B MoE agent model built on Agents-A1, was introduced with two key capabilities: agentic orchestration and agentic verification. These capabilities extend the model's ability to organize work and revise its approach during execution, offering a practical path from long-horizon task execution to collaborative, verifiable, and inspectable scientific exploration.
Adaptive Agentic Execution: Orchestration and Verification
Sage-A1.5 operates along two complementary dimensions. Capability coverage spans the range of domains and tasks an agent can address, a dimension expanded by domain-specific training. Capability composition concerns how problem-solving capabilities are organized and adapted during execution, a dimension developed through agentic orchestration and agentic verification. Agentic orchestration determines how work is decomposed, delegated, and integrated across agent executions. Agentic verification determines when to reassess intermediate judgments and exploration strategies, using findings to revise subsequent actions. Together, these capabilities make execution organization and feedback-driven adjustment explicit targets of model training alongside domain-specific problem solving.
Agentic orchestration extends single-agent execution through task delegation and parallel collaboration among subagents with isolated contexts. The model learns to decide whether and how to decompose work, which subtasks to delegate, what context to provide, and how to coordinate dependencies and integrate results. Two complementary patterns of parallel agent execution are distinguished. Exploratory composition assigns multiple agents different hypotheses or solution approaches to the same question; their outputs provide alternative candidates and supporting evidence that the main agent compares to narrow the investigation and decide which directions to pursue. Constructive composition decomposes an overall task into complementary subproblems and assigns them to agents for parallel work; each output contributes a distinct part of the solution that the main agent integrates, coordinating dependencies and checking coverage and consistency. Complex tasks dynamically interleave both patterns as emerging evidence informs task decomposition, exploration, and integration.
The orchestration protocol involves the main agent investigating directly or delegating subtasks, specifying their scope and context. Subagents search and read webpages in separate contexts, advancing their model calls and tool interactions independently within a concurrent batch. Once the batch completes, the main agent uses the findings to synthesize an answer or organize further investigation. Building on the model's ability to critique its own execution, each subagent returns its result together with a self-assessment, supporting evidence, and remaining issues. This feedback helps the main agent identify gaps and conflicts and provides signals for trajectory selection. Task construction extends from relation-chain methods into constraint-centric graphs, where an answer entity is selected, directly related entities providing different relation constraints are identified, and chains are extended from each, forming a starlike tree graph around the shared answer. Branch descriptions and evidence are combined into one question with identifying details masked, allowing the main agent to delegate different constraints for parallel investigation.
Agentic verification strengthens long-horizon reliability through progress checking, evidence validation, and reflective strategy revision. In long-horizon research tasks, agents must gradually build chains of evidence through multiple rounds of retrieval and reasoning. When an exploration path repeatedly fails to yield useful information, the model may continue searching around the same hypothesis or use unverified assumptions as premises for subsequent reasoning, leading to ineffective exploration and accumulating errors. Agentic verification enables the model to adaptively recognize when verification is needed, examine existing evidence and exploration strategies within its own reasoning, and adjust subsequent actions accordingly. When progress stalls or an important intermediate conclusion is reached, the model reviews facts confirmed by evidence, examines earlier judgments for unsupported inferences, identifies unresolved information gaps, and determines which exploration paths have proved ineffective and how to proceed. The central idea is to make verification a learned capability of the model itself: the model adaptively determines when to verify, what to verify, and how to continue based on the findings, rather than having verification triggered and executed by an external harness according to predefined rules. This mechanism enables the model to proactively revise its judgments, gather additional evidence, or change its search direction, improving its ability to self-correct during long-horizon tasks.
To train agentic verification, complete research trajectories containing intermediate reflections and the actions that follow them are constructed. During data construction, an execution model conducts research through live search and webpage reading, while a supervisor model examines the exploration process and introduces verification at points involving repetitive searches, sustained lack of progress, or key conclusions requiring confirmation. Each verification relies only on the evidence available at that point, distinguishes confirmed facts from unverified hypotheses, and proposes a concrete direction for further investigation. The execution model then continues exploring based on the verification findings, producing coherent training examples that connect problem recognition, evidence assessment, and action adjustment. After data cleaning, external supervision prompts and control messages are removed, and reflections are incorporated directly into the model's reasoning text. This allows the model to learn when to verify, what to examine, and how verification should inform subsequent actions. A conditioning marker during training enables the capability on demand: adding "Verification protocol: enabled" to the system prompt activates agentic verification. Once enabled, the model performs verification within its regular reasoning and tool-use process without requiring an additional verification tool, an online supervisor model, or a harness-implemented procedure. Context management uses intermediate verification points as natural boundaries for organizing exploration history; segments between the two most recent verification points may be consolidated into concise stage records, preserving key evidence, rejected hypotheses, unresolved questions, and the current task state.
Three-Stage Training Pipeline Unifying Domains and Modes
Sage-A1.5's unified training pipeline combines multi-domain supervised fine-tuning with mixed-reinforcement learning, fusing capabilities across domains, reasoning modes, and agent harnesses into a single agentic model. The pipeline proceeds in three stages. First, diverse domain data including coding, long-context reasoning, deep research, scientific reasoning, and co-working are used to train environment construction, trajectory selection, staged supervised fine-tuning, and reinforcement learning, building multi-domain teachers. Second, these teachers undergo unified outcome-process post-training to produce the final unified model, which features agent orchestration for collaborative task solving, iterative agentic verification for reasoning reliability, and local deployment for ensuring data privacy, access control, and offline support. Third, recursive task evolution for long-context reasoning constructs questions from LongBlocks documents spanning 16K–128K tokens using 24 fine-grained task categories, applies atomic capabilities and compositional enhancement along ten difficulty axes, and uses feedback-guided recursive evolution with reference-model performance as a selection signal to generate hard QA pairs and high-quality reasoning trajectories.
Post-training on engineering tasks includes MLE and SWE trajectories. MLE tasks are constructed from historical Kaggle competitions via an automated multi-agent pipeline that converts competitions into executable tasks with public workspaces, hidden test sets, and a unified evaluation interface. The interaction harness allows the model to autonomously refine solutions based on execution feedback, organizing candidate code and execution results into an experiment tree that records solution revisions. Trajectory collection and selection truncates persistently degrading tails while preserving useful exploration and failure recovery. Staged supervised fine-tuning first establishes code revision and tool-use capabilities with general coding and tool-interaction trajectories, then combines MLE trajectories with experimental trajectories from literature-derived scientific tasks to train solution design, result analysis, and sustained optimization. Windowed interaction reinforcement learning groups interactions between successive experiment results into execution windows with shared outcome-based rewards, and asynchronous policy optimization corrects policy differences using token probability ratios.
SWE tasks involve repository-level software-engineering issue resolution. Each task provides a repository snapshot and a natural-language issue description, and the agent interacts with the repository to produce a patch. The patch is evaluated against held-out tests: designated failing tests must pass, while previously passing tests must remain passing. Task sources include SWE-rebench-V2 and in-house constructed tasks, with environment configurations adapted to the cluster. Training corpus includes successful trajectories and a filtered subset of unresolved attempts meeting criteria for non-empty patches, substantive task-directed interaction, well-formed structure, passing leakage checks, and additional harness coverage. Leakage control restricts access to reference fixes and held-out tests during agent execution, forces outbound traffic through a cluster proxy, and redirects the proxy endpoint to a closed local port during rollout. Trajectory collection uses two agent harnesses, mini-swe-agent and Claude Code, and retained trajectories must pass audit verdicts and not overlap evaluation task splits.
Scientific research enhancement follows a three-stage pipeline: scientific problem selection, scientific research harness optimization, and trajectory distillation and training. Problem selection identifies high-quality scientific problems based on difficulty, task type, and tool benefit, prioritizing medium-to-high difficulty problems requiring substantial reasoning or research-oriented exploration. Harness optimization improves a scientific research harness equipped with search, visit, code, and scholar tools, introduces an isolated code execution sandbox, and implements a tool-result replay mechanism for search, visit, and scholar to reduce redundant calls and improve reproducibility. Trajectory distillation and training uses a strong model with the optimized harness to distill long-horizon trajectories on selected problems, filters along solution correctness, reasoning depth, and tool-call success rate, and performs extended multi-epoch token-level supervised fine-tuning on Qwen3.5-35B-A3B using filtered demonstrations that also retain trajectories involving recoverable tool failures.
Benchmarks and Capability Assessment
Sage-A1.5's capabilities are evaluated across a broad suite of benchmarks. Despite having only 35B parameters, Sage-A1.5 outperforms substantially larger models including Kimi-K3, GLM-5.3, and GPT-5.6-solve on FrontierScience-Olympiad, CMT, PinchBench, and IFBench. On SWE-Bench Pro and BrowseComp, Sage-A1.5 remains competitive, and on all twelve benchmarks it matches or surpasses the equally sized Agents-A1 and Ornith-1.5. These results demonstrate that adaptive capability composition delivers measurable gains over pure capability coverage, and that a 35B MoE architecture can compete with much larger models when execution structure is optimized.
The benchmark suite spans scientific reasoning, tool use, instruction following, multi-step search, software engineering, and economically valuable real-world tasks. FrontierScience-Olympiad tests graduate-level scientific knowledge across disciplines. CMT and PinchBench evaluate code modification and patch generation. IFBench assesses instruction following with fine-grained categories. SWE-Bench Pro measures repository-level issue resolution. BrowseComp measures web-based reasoning. Economically valuable tasks are drawn from Agents' Last Exam, which covers 55 subdomains across 13 industry clusters. The broad coverage ensures that gains from agentic orchestration and verification are not confined to a narrow capability subset.
Open Scientific Discovery via Sage-A1.5
Sage-A1.5 extends agentic capabilities beyond answering predefined questions to conducting exploratory research. This third improvement, open scientific discovery, connects scientific inquiry with interactive analysis and inspectable research artifacts. The system showcases these capabilities through complicated agentic tasks and multidisciplinary scientific applications, offering the community an accessible foundation for open scientific discovery.
Foundational capabilities are demonstrated first. Agentic orchestration is shown splitting a broad research objective into complementary parallel assignments, assessing coverage and consistency of findings, and integrating them into a report answering the original question. Agentic verification is shown recognizing unproductive search paths, reviewing evidence and assumptions, and adjusting subsequent actions to pursue alternative approaches. Local automated deployment is shown ensuring data privacy, access control, and offline support for scientific workflows.
Scientific information integration and long-horizon reasoning are illustrated through case studies. Sage-A1.5 integrates evidence from multiple sources into coherent research reports, tracks claims to their source documents, and identifies open questions warranting further investigation. The system can maintain reasoning continuity across hundreds of token interactions, filter redundant retrieval, and synthesize conclusions supported by evidence from diverse sources including web search, academic literature, and computational verification.
From scientific co-working to end-to-end auto-research, Sage-A1.5 supports multiple modes. In scientific co-working mode, the model assists human researchers by suggesting research directions, summarizing relevant literature, and designing computational experiments. In end-to-end auto-research mode, the model independently conceives a problem, conducts a literature review, formulates hypotheses, designs and runs computational tests, interprets results, and produces a written report with figures and tables. Problem-solving aligned by scientists mode lets scientists specify constraints and objectives, after which the model generates a research plan, tracks progress, and provides intermediate reports for scientist review and course correction.
Case studies span disciplines. In computational biology, Sage-A1.5 analyzed protein-ligand binding data, cross-referenced findings with recent literature, and proposed mutation strategies with predicted affinity improvements. In climate modeling, the model integrated observational data, ran parameter sweep simulations, and generated a summary of temperature trend attributions consistent with IPCC assessment principles. In materials science, the model screened literature for synthesis protocols, evaluated computed formation energies, and recommended candidate compositions for targeted experimental validation. Each case study produced inspectable research artifacts including intermediate reasoning traces, evidence provenance graphs, and structured reports with claim-evidence pairs.
Limitations and Future Directions
Sage-A1.5's operating context is primarily long-horizon text-based tasks; complex multimodal reasoning beyond image and document inputs remains an area for future investigation. The model's capacity to maintain consistent reasoning across very long contexts is constrained by the attention mechanism of the underlying MoE architecture, and further context compression improvements would extend viable horizon length. The current verification mechanism relies on evidence available within the model's reasoning trace; entirely external evidence stores not accessible through tool use are not integrated. Additionally, the training pipeline's dependence on strong teacher models for trajectory distillation means that capabilities are bounded by the teacher's own generalization capacity, and expanding the teacher pool would be a natural next step. The authors also note that computational cost grows with the number of parallel subagents in orchestration, and more efficient delegation policies that minimize unnecessary parallel work would improve scalability.
Future work includes integrating vision and audio capabilities, developing more sophisticated context compression and retrieval-augmented generation for very long horizons, extending verification to incorporate external fact-checking APIs, and adapting the orchestration framework to robotic and simulation domains. The open scientific discovery pathway is intended as a starting point, and community contributions of domain-specific problem sets, evaluation suites, and artifact formats are encouraged.
Practical Path from Task Execution to Scientific Exploration
For researchers and engineers looking to build capable long-horizon agents, Sage-A1.5 offers several practical takeaways. First, orchestration and verification are learned model capabilities, not fixed external rules; training data must include examples of task decomposition, parallel collaboration, evidence-driven self-correction, and strategy revision. Second, a unified training pipeline that spans domains, reasoning modes, and harnesses produces more robust generalist agents than domain-specific fine-tuning alone. Third, inspectable research artifacts—intermediate reasoning traces, evidence provenance graphs, and structured reports with claim-evidence pairs—are essential for verifiable scientific exploration and should be built into the agent's output format from the outset. Fourth, the conditioning markers that enable capabilities on demand (e.g., "Verification protocol: enabled") provide a lightweight way to activate specialized behaviors without deploying separate supervisor models or harness procedures. Finally, the recursive task evolution method for generating hard QA pairs and high-quality reasoning trajectories is a general technique that can be applied to any long-context reasoning domain.
Sage-A1.5 demonstrates that a 35B MoE agent model, when trained with adaptive capability composition, can outperform much larger models on a broad suite of benchmarks and support multidisciplinary scientific exploration. The shift from isolated task solving toward long-horizon scientific research represents a practical step toward agents that can not only answer questions but also conduct verifiable, inspectable research. The community now has a foundation for open scientific discovery, offering a practical path from long-horizon task execution to collaborative, verifiable, and inspectable scientific exploration.
Read the paper on arXiv