PaperlayerView exact revision on arXiv

Source-linked AI summary

Apodex 1.1: Scaling Agentic Intelligence for Complex Work

Apodex Team, B. An, B. Li, B. Wang, B. Zhang, B. L. Wang, C. Feng, C. Wei, C. Xue, C. Zhang, D. Ng, D. Ye, E. Min, F. Chen, F. Liu, F. Yang, F. Ye, H. Xu, H. Yang, H. Ye, H. Zhang, H. Zhao, J. Li, J. Lin, J. Xia, K. Jin, K. Wang, K. Yang, L. Bing, L. Lei, L. Su, Le. Wang, Lu. Wang, N. Wang, Q. Ren, Q. Yang, R. Li, S. Bai, S. Du, S. Li, S. Lin, S. Nie, S. Wang, S. Zhang, S. Z. Wang, Ta. Q. Fang, Ti. Q. Fang, W. Fang, W. Li, W. Zhang, X. Chen, X. Li, X. Tang, X. Wang, X. Xu, X. Zhang, X. Q. Wang, X. Y. Wang, Y. Deng, Y. Gao, Y. Hu, Y. Li, Y. Sui, Y. Wang, Y. Xiao, Y. Zhang, Z. Chen, Z. Cheng, Z. Feng, Z. Liang, Z. Zhang

arXiv:2608.23283v1cs.AIcs.CLcs.LG

TL;DR

Complex work demands sustained, verifiable interaction with files, information, code, state, and failures beyond answer generation. Apodex 1.1 develops this working capability through environment scaling and agentic coordination, connected by a shared execution stack and trajectory-based training. It reaches the leading performance band across broad task families, including at a compact 35B-parameter scale, while component-level causal attribution remains limited.

  • Problem

    Complex tasks require sustained tool interaction, state maintenance, failure recovery, and verifiable delivery, but reasoning and answer quality alone do not establish reliable work.

  • Method

    Apodex 1.1 scales executable file, search, and code environments alongside trained delegation, integration, replanning, and recovery, using a shared harness and Agent Team.

  • Results

    Apodex 1.1 reaches the performance band of selected frontier systems across representative professional, financial, and scientific tasks, while its 35B mini improves over Apodex 1.0 mini and retains Agent Team gains.

  • Takeaways & Limitations

    The paper grounds agentic intelligence in completed, verifiable work over time and establishes a development path toward a Heavy-Duty Solver.

  • Takeaways & Limitations

    The HDS6 comparisons are within-panel system comparisons and do not isolate the causal contribution of individual system components.

Abstract

from arXiv · show

General-purpose language models can reason and synthesize knowledge, but complex work also requires sustained interaction with files, information sources, and executable code, together with state maintenance, failure recovery, and verifiable delivery. We call this \emph{working capability}: sustained, verifiable progress toward a real-world objective. Apodex 1.1 develops this capability along two complementary dimensions. \emph{Environment Scaling} expands the diversity and verifiability of executable file, search, and code environments, while \emph{Agentic Coordination Scaling} trains agents to decompose long-horizon tasks, delegate parallel work, integrate asynchronous results, and replan. A shared execution harness and AgentOS maintain task state and provenance across tools and agents, and training turns environment trajectories and coordination traces into reliable behavior. Across complex professional work, finance, scientific research, mathematics, coding, and search, Apodex 1.1 reaches the leading performance band despite using a substantially smaller model than many frontier systems. The 35B-parameter Apodex 1.1 Mini further retains strong working capability in a locally deployable form. These results ground agentic intelligence in useful, verifiable work completed over time and advance our goal of building a \emph{Heavy-Duty Solver} for ambitious, long-running tasks.

1 Introduction

Apodex 1.1 treats complex work as sustained, verifiable progress rather than isolated reasoning, and develops it through environment scaling, agentic coordination, shared execution infrastructure, and trajectory-based training.

  • Motivation: Complex work requires sustained interaction with files, information sources, code, state, failures, and delivery checks beyond producing a correct answer.The paper identifies the limiting capability as turning reasoning into sustained work inside a changing environment.
  • Working capability: Working capability is defined as making useful progress over long horizons while acting through tools, revising plans, recovering from failure, and satisfying a delivery contract.The unit of capability is completed, inspectable work rather than an isolated response.
  • Results: Apodex 1.1 reaches the strongest values shown in FrontierFinance and FrontierScience-Research comparisons while remaining competitive across professional work, reasoning, search, mathematics, and coding.The 35B-parameter Apodex 1.1 mini reaches the performance band of selected frontier systems and improves over Apodex 1.0 mini on overlapping tasks.
  • Environment Scaling: Environment Scaling expands diverse, faithful, and verifiable file, search, and code worlds to provide executable trajectories for learning.These environments expose actions, observations, file changes, code outputs, recovery decisions, and artifacts as part of the learning distribution.
  • Agentic Coordination Scaling: Agentic Coordination Scaling trains delegation, staged integration, shared-state revision, and dynamic replanning, while Agent Team realizes these behaviors asynchronously at runtime.The approach organizes long-horizon work across agents and task branches rather than treating coordination as a separate wrapper.
  • Execution infrastructure: A common execution harness connects executable environments with persistent workspace, branch state, provenance, replay, and verification, with AgentOS supplying the runtime.The same infrastructure supports stateful execution and verifiable artifact delivery.
  • Training and evaluation: Unified SFT and agentic RL train one policy on environment trajectories and coordination traces to improve execution, recovery, delivery, and coordination.Evaluation spans professional work, science, finance, mathematics, coding, search, and end-to-end cases.

2 Core Insights and Design Principles

The paper defines agentic intelligence around completed, verifiable work in changing environments. Its design combines an executable task contract, persistent state and provenance, coordinated branching, and training over environment and coordination trajectories.

  • Core observation: High-quality reasoning is necessary but insufficient when complex work requires sustained progress through a changing task.The target capability includes preserving state, recovering from failures, organizing computation, and meeting checkable delivery objectives.
  • Completed work: The task contract E specifies the workspace, initial state, objective, actions, transitions, observations, budgets, delivery contract, and verifier.The transition operator may be stochastic, so replay must preserve exogenous state, tool versions, and random seeds.
  • Completed work: The delivery contract defines what must be delivered, which constraints apply, and how completion is judged, while q is the normalized task objective.Clarifications that do not materially change the outcome remain within the same task contract.
  • Completed work: The execution trace records the initial request, actions, observations, and asynchronous interventions across the realized horizon.Tool failures, timeouts, verifier results, and subagent reports are runtime responses rather than user inputs.
  • Completed work: Success requires both a useful result and a defensible path from input to delivery, because the terminal verifier evaluates contract-specific outcomes rather than prose alone.Verifier clauses may use tests, recomputation, source alignment, structured rubrics, or human review.
  • Environment Scaling: Environment Scaling expands executable task-contract distributions across file, search, and code worlds, while preserving valid transitions and evaluable outcomes.These worlds teach artifact handling, evidence reconciliation, executable transformation, testing, and recovery.
  • Agentic coordination: Agent Team continuously moves results, decisions, and feedback through active work so the lead can revise priorities, terminate obsolete branches, and create new workstreams.A slow or failed branch does not erase completed work elsewhere.
  • Training: Unified SFT and agentic RL convert executable environment trajectories and coordination traces into common task-execution and coordination behavior.Training covers decomposition, delegation, staged return, integration, replanning, and recovery.

3 Apodex 1.1

Apodex 1.1 is a common policy and execution stack for long-horizon work, combining executable environments, learned coordination, persistent state, replay, and completion checks.

  • System design: Apodex 1.1 trains one policy to reason, search, manipulate files, execute code, recover from failures, coordinate parallel work, and deliver verifiable artifacts.These behaviors are developed as integrated capabilities rather than disconnected specialist modes.
  • Scaling dimensions: Environment Scaling expands the executable file, search, and code worlds in which the policy learns and acts.Agentic Coordination Scaling expands how work is decomposed, delegated, integrated, verified, and revised across agents and time.
  • Execution stack: A common harness binds both scaling dimensions to persistent state, typed observations, replay, and completion checks, with AgentOS providing the execution substrate.The architecture converts both trajectory families into sustained runtime execution.

3.1 Environment Scaling

Environment Scaling treats executable worlds as a primary scaling surface, expanding file, search, and code situations with diverse state transitions, evidence requirements, and verifiable delivery conditions.

  • Motivation: Complex work fails when models cannot locate authoritative files, preserve state, recover from execution errors, reconcile evidence, or verify completed artifacts.These failures determine whether reasoning becomes reliable long-horizon work.
  • Scaling principle: Environment Scaling increases the coverage, fidelity, and structural depth of executable situations rather than merely adding tools or prompts.The policy learns from increasingly diverse state transitions and observable consequences.
  • Task contract: The environment task contract spans initial workspaces and objectives, action spaces, transitions and observations, resource budgets, and delivery-verification pairs.Apodex develops this distribution across file, search, and code worlds under common construction, verification, and replay requirements.
  • File worlds: File worlds distribute facts across heterogeneous workspaces, requiring authority discovery, relationship reconstruction, transformation, and independently recoverable evidence.Coverage spans 33 domains, 318 occupations, and 1,208 deliverable clusters, while structural depth varies independently of file count.
  • Search worlds: Search worlds model research as discovery, acquisition, and evidence synthesis, with gold outputs containing source sets, claim-to-evidence alignment, and explicit uncertainty.The agent must formulate queries, triage sources, follow references, reconcile evidence, and stop when support is sufficient.
  • Code worlds: Code worlds require agents to change repositories, dependencies, and processes, then verify resulting executable states through sandboxed tests and hardened graders.Synthesized tasks are admitted only after checking that reward cannot be obtained without completing the task.
  • Capability-driven construction: New worlds are allocated to capability gaps exposed by failed trajectories, then re-enter family-specific verification before contributing training signal.Model-based diagnosis selects exploration areas, while environment verification determines which worlds are trustworthy for training.
  • Difficulty calibration: Acquisition-heavy difficulty tracks plausible candidates, load-bearing evidence transitions, and tool-call budgets rather than raw file or page counts.The resulting acquisition-pressure coordinate is a first-order measure, not a universal difficulty metric; code worlds require different calibration coordinates.

3.2 Apodex Agent Team 1.1: Interactive Self-Organizing Teams

Apodex Agent Team 1.1 makes coordination an inspectable, revisable task-state process rather than a one-shot internal decision. It combines adaptive delegation, user intervention, asymmetric verification, and evidence-grounded synthesis for long-horizon work.

  • The lead agent decomposes problems globally and constructs specialized subagents on demand, so team structure is induced by each task.
  • The persistent task board records bounded objectives, dependencies, owners, resolution states, and returned evidence or artifacts before delegation.
  • Externalized coordination lets results unlock dependent tasks, isolates invalidated descendants, and keeps independent branches progressing while requiring continual reconciliation with inspectable state.
  • Execution-time user messages can clarify requirements, correct facts, change priorities, revise methods, or add constraints while preserving completed work that remains causally valid.
  • Verification narrows each reviewer’s task to attacking a claim, its evidence, and its delivery constraint instead of reproducing the entire solution.
  • Adaptive Max Team Effort assigns additional independent investigations to weak, contested, or load-bearing claims rather than scaling every subproblem uniformly.
  • Evidence-graph construction reconciles branch results before drafting, marking claims as corroborated, disputed, rejected, or unresolved for the writer agent.

3.3 AgentOS: An Execution Substrate for Long-Horizon Work

AgentOS provides the persistent execution substrate for single- and multi-agent long-horizon work. It manages workspace, provenance, coordination state, interventions, context pressure, deadlines, and partial recovery under explicit runtime contracts.

  • AgentOS maintains persistent workspace and tool state, exposes runtime events as observations, preserves progress, and controls conversion of intermediate artifacts into deliverables.
  • The workspace tracks file state, retrieved evidence, executable state and logs, artifact indexing, dependency graphs, and optional runtime control state.
  • Runs separate read-only inputs, writable workspace, and output namespaces, with optional read-only access to a durable cross-run document library.
  • Workspace visibility depends on isolation: subagents use private inspectable worktrees or shared disjoint paths, while inputs remain read-only and outputs shared.
  • The task board remains outside model history, separates coordinator-owned semantics from runtime job status, and is re-injected after context compaction.
  • Planning mode blocks delegation, file mutation, and finalization before planning ends, while execution permits board revision and requires active items to resolve or cancel.
  • AgentOS supports execution lifecycle tracking, asynchronous interventions, tiered context compaction, soft-deadline consolidation, bounded finalization, and preservation of completed work after interruption.
  • On soft-deadline exhaustion, retries and tool calls respect remaining time, and a tool-free finalization call recovers a useful partial result before hard cancellation.

3.4 Training

Training represents delegation, staged return, synthesis, branch revision, and recovery directly in trajectories rather than treating them as inference-time wrapper properties. It combines broad supervised initialization with reinforcement learning over heterogeneous long-horizon environments.

  • Training trajectories encode delegation, staged return, synthesis, branch revision, and recovery as behaviors the policy must strengthen.
  • The SFT mixture unifies reasoning, tools, search, files, coding, mathematics, scientific and financial reasoning, professional delivery, and multi-agent coordination.
  • Filtering removes invalid interactions, inconsistent state, ignored observations, and incomplete delivery, while rubric-based selection retains task-relevant high-quality demonstrations.
  • Model-soup merging combines domain-specialized SFT variants into a unified behavioral initialization for later optimization.
  • Reinforcement learning targets sustained progress across file, code, search, and coordination environments with varied costs, lengths, and failure modes.
  • PIVOT-RL localizes optimization at consequential decisions where trajectories become unproductive, weakly evidenced, tool-misusing, or resistant to revision.
  • Localized continuation tasks preserve useful prefixes, restore executable state when needed, and combine corrective training with unhinted full-task solving.
  • Search, knowledge, and science evaluations improve consistently as reinforcement learning compute scales, while completed trajectories enter optimization asynchronously.

4 Evaluation

Apodex 1.1 is evaluated across public agentic, knowledge-intensive, and model-scale-efficiency benchmarks using ReAct and Agent Team execution modes. Agent Team adds organized computation and achieves leading results across multiple task families, while the 35B Mini reaches frontier performance bands.

  • Evaluation setup: The evaluation spans public benchmarks, internal structured search, end-to-end research delivery, and execution-process reliability.The progression moves from outcomes to capability-specific evidence and then to the reliability of the work producing them.
  • Main results: Apodex 1.1 Mini uses 35B parameters and reaches the performance band of selected frontier systems across representative work, finance, and scientific-research tasks.The comparison is stated as a performance-band result because parameter counts for several proprietary reference systems are not public.
  • Evaluation setup: Apodex 1.1 is evaluated with ReAct and Agent Team across public agentic and knowledge-intensive benchmarks.ReAct minimizes orchestration, whereas Agent Team dynamically spawns subagents to parallelize or specialize work, integrate results, and revise shared plans.
  • Main results: Agent Team converts additional coordinated computation into further gains, including the strongest values shown in FrontierFinance and FrontierScience-Research comparisons.The system remains competitive on GDPVal, APEX-Agents, HLE, and DeepSearchQA.
  • Main results: Agent Team raises Mini scores from ReAct by 10.2 points on FrontierFinance, 6.7 on FrontierScience-Research, and 3.5 on APEX-Agents.The corresponding Agent Team scores are 50.2, 51.7, and 27.7, respectively.

4.3 Capability Analysis

Capability analyses cover professional work, finance, scientific research, reasoning and search, mathematics, and coding. Across these domains, Agent Team improves performance over ReAct, while Apodex 1.1 shows substantial gains over earlier versions on several evaluations.

  • Professional Work: GDPVal reaches 78.8 with Agent Team and 69.5 with ReAct, a 9.3-point gain on professional artifact quality.GDPVal covers 44 occupations across nine industries and reports win rate.
  • Professional Work: APEX-Agents reaches 38.5 with Agent Team and 34.4 with ReAct, compared with 16.5 for Apodex 1.0.The benchmark tests 480 long-horizon, cross-application tasks in 33 data-rich worlds.
  • Finance: 54.3 on FrontierFinance with Agent Team exceeds the ReAct score by 5.6 points.The benchmark covers investment workflows including discovery, research, modeling, earnings analysis, and portfolio monitoring.
  • Scientific Research: 63.3% on FrontierScience-Research with Agent Team improves over 55.0% with ReAct and 28.3% for Apodex 1.0.The benchmark comprises 60 research-level physics, chemistry, and biology tasks.
  • Reasoning and Search: Agent Team reaches 56.1 on HLE and 92.4 F1 on DeepSearchQA, compared with 53.2 and 88.2 for ReAct.These benchmarks test broad expert-level reasoning and multi-step evidence seeking and synthesis.
  • Mathematical Reasoning: On the evaluated competition mathematics sets, Apodex 1.1 with Agent Team exceeds the stated reference threshold on all three sets, including IMO gold-medal cutoffs.MathArena-derived scores rise over Apodex 1.0 from 12.5 to 36.5 on IMO 2025, 13.0 to 30.5 on IMO 2026, and 5.8 to 26.5 on USAMO 2026.
  • Coding: Apodex 1.1 reaches 70.8 on Terminal-Bench 2.1 and 77.7 on SWE-bench Verified.These evaluations test executable changes, feedback from commands or tests, and repository-level issue resolution.

4.4 Internal Evaluation

Internal evaluations target structured evidence acquisition and end-to-end scientific workflows with verifiable deliverables. Agent Team achieves the best FrontierSearchBench comparison result, while FrontierResearchBench shows that full scientific-workflow completion remains difficult.

  • FrontierSearchBench: FrontierSearchBench evaluates 41 verifiable deep-search tasks with structured deliverables assembled and reconciled across many sources.Scoring extracts claims, aligns them to frozen ground-truth dimensions, and penalizes incorrect assertions.
  • FrontierSearchBench: The benchmark reports an unweighted mean on a 100-point scale across 41 task contracts, with a mathematical range of [-100, 100].Negative aggregate scores are possible when hard-negative penalties exceed positive credit.
  • FrontierSearchBench: Agent Team lifts FrontierSearchBench’s average score past the strongest proprietary reference shown and roughly halves the share of tasks earning no credit relative to ReAct.It also keeps incorrect assertions rare.
  • FrontierResearchBench: FrontierResearchBench contains 97 executable scientific tasks spanning multiple disciplines and evaluates end-to-end delivery rather than isolated question answering.Custom graders check required files, numerical results, formats, executable outputs, and consistency across artifacts.
  • FrontierResearchBench: 12.4% is Apodex 1.1 Agent Team’s Pass Rate on the high-difficulty FrontierResearchBench workflows.Even GPT-5.6-Sol with Codex and Grok-4.6 with Claude Code receive full credit on only 20.6% of tasks.

4.5 Heavy-Duty Solver (HDS6) Analysis

HDS6 evaluates whether successful outputs are supported by coherent, evidence-grounded, recoverable processes. Its comparisons show improvements in specific process capabilities, while the setup limits causal attribution and direct cross-panel comparison.

  • HDS6 evaluates six process capabilities to determine whether task outcomes reflect coherent, grounded, and recoverable work.The framework complements outcome scores with process verification.
  • Hundreds-step asynchronous multi-agent trajectories are stitched into ordered logs and scored through role-specialized mapping, judging, and review.
  • Deep Discover, Deep Solve, and Deep Research use different task families, model scales, execution modes, and aggregation protocols, so scores should be compared within panels.Deep Discover v1.0 aggregates eight runs, whereas v1.1 uses one run.
  • +1.3 for Initial Decomposition and +0.8 for Final Verification are the two largest single-item deltas.Evidence fidelity and hypothesis management also improve across constituent rubric items.
  • HDS6 localizes where combined model, environment, coordination, and harness changes appear in the work process but does not isolate individual component effects.
  • Three end-to-end cases cover clinical survival analysis, molecular docking, and electrochemical corrosion analysis.The cases document inputs, work division, tool execution, intermediate checks, and delivered artifacts.

5 Related Work

Related work develops agentic systems through executable environments, open-ended search, scientific workflows, multi-agent coordination, verification, and reinforcement learning. These strands motivate Apodex 1.1’s emphasis on sustained, tool-mediated work.

  • General-purpose frontier models increasingly combine long-context reasoning, coding, tool use, information seeking, and sustained task execution.
  • Executable-environment research connects reasoning and action with reproducible websites, enterprise workflows, and operating-system tasks.
  • Deep-research agents extend reasoning-and-acting toward open-ended information seeking, evidence reconciliation, and report synthesis.
  • Scientific agents combine domain tools with code execution, experimentation, visualization, writing, and review in automated research loops.
  • Multi-agent systems organize programmable conversations, specialized roles, collaborative groups, debate, and test-time scaling.
  • Verification research studies iterative self-feedback, stored verbal feedback, targeted verification questions, and process-level reward assessment.
  • Recent work treats environment construction as a scaling dimension for agent learning, including synthesized tool-use and evolving real-world environments.

6 Conclusion

Apodex 1.1 frames agentic intelligence as sustained progress through changing tasks, supported by executable environments, coordinated agents, persistent state, and unified training. Its evaluations show broad capability and compact-scale efficiency, motivating the Heavy-Duty Solver direction.

  • Apodex 1.1 defines agentic capability by sustaining progress through changing tasks, using tools and evidence, recovering from failure, and delivering verifiable results.
  • Environment Scaling broadens executable file, search, and code worlds, while Agentic Coordination Scaling broadens decomposition, delegation, integration, and reorganization across agents and time.A common harness and AgentOS provide persistent execution state, and unified training converts trajectories into a stronger working policy.
  • Evaluations span scientific research, professional work, mathematics, search, coding, and internal long-horizon tasks, while ReAct–Agent Team comparisons estimate coordination lift.HDS6 and end-to-end cases examine process coherence, evidence, repair, and tool use.
  • The 35B Apodex 1.1 mini reaches the performance band of selected frontier systems on representative professional, financial, and scientific tasks.It improves over Apodex 1.0 mini and retains substantial Agent Team gains on overlapping evaluations.
  • The longer-term Heavy-Duty Solver goal is responsibility for increasingly ambitious, long-running, and verifiable work.The proposed path emphasizes broader executable environments, longer-horizon coordination, hierarchical credit assignment, and feedback from failures into training and evaluation.

A Case Studies

The case studies show Apodex agents producing checkable artifacts across molecular simulation, image-based apoptosis quantification, and pig RNA-seq analysis. They also expose practical boundaries involving unavailable software, definitional choices, statistical significance, and correlation-based interpretation.

  • Case studies: Three recorded runs required finished, checkable artifacts rather than answers: a protein simulation package, an apoptosis-analysis report, and a pig gene-network analysis.The cases used raw files and specified deliverables, with agent traces documenting execution and delivery.
  • Case 1: Protein simulation: The molecular-simulation case delivered 22 corrected files and cross-checked six chain topologies with 0 mismatches.The package included repaired structures, regenerated inventory data, and topology checks despite the simulation engine being unavailable.
  • Case 2: Apoptosis quantification: The apoptosis arbitration selected shared-ROI means because per-mask means changed substantially with thresholds and could reverse the group comparison.Under the selected threshold, the reported ratio was computed from identically processed TUNEL and DAPI channel means.
  • Case 2: Apoptosis quantification: The apoptosis case found a 2.4× control-to-experimental TUNEL/DAPI ratio difference, but Welch two-sample P = 0.0593 and n = 3 per group prevented a significance claim.The analysis used shared tissue-ROI channel means and reported the effect as large but non-significant.
  • Case 3: Pig RNA-seq WGCNA: The WGCNA case produced 437 ranked hub-candidate records across five phenotypes, while explicitly treating them as candidates rather than causal or regulatory genes.Each phenotype mapped to a different non-grey module at FDR < 10^-16, but the result remains correlation-based.
  • Case 3: Pig RNA-seq WGCNA: The WGCNA conclusions are bounded by n = 50, unavailable batch information, non-directional correlation, and sensitivity of module partitions to analysis parameters.The durable outputs are the module–trait associations and candidate ranking, not the module colours themselves.

B Contributors

The contributor list is presented in alphabetical order by given-name initials.

  • Contributors are ordered alphabetically by their given-name initials.
Loading 2608.23283v1…