Summary of "Apodex 1.1: Scaling Agentic Intelligence for Complex Work"

Summary (Overview)

• Core Contribution: Apodex 1.1 introduces a general-purpose model and execution system that scales "working capability"—the ability to make sustained, verifiable progress on complex, long-horizon tasks—through two complementary scaling dimensions: Environment Scaling (expanding executable file, search, and code worlds) and Agentic Coordination Scaling (training decomposition, delegation, and replanning behaviors).

• Key Architecture: A unified policy and execution stack called AgentOS provides persistent workspace state, a task board for multi-agent coordination, controlled artifact delivery, and mechanisms for failure recovery, context management, and user intervention.

• Performance Results: Apodex 1.1 with Agent Team achieves leading performance on FrontierFinance (54.3) and FrontierScience-Research (63.3), and competitive results on GDPVal (78.8), APEX-Agents (38.5), Humanity's Last Exam (56.1), and DeepSearchQA (92.4 F1).

• Model-Scale Efficiency: The 35B-parameter Apodex 1.1 Mini reaches the performance band of selected frontier systems, with ReAct scores of 40.0 (FrontierFinance), 45.0 (FrontierScience-Research), and 24.2 (APEX-Agents), improving to 50.2, 51.7, and 27.7 with Agent Team.

• Training Approach: Uses a unified SFT mixture followed by PIVOT-RL (localized trajectory optimization at consequential decision points) and asynchronous RL, converting environment trajectories and coordination traces into reliable model behavior.

Introduction and Theoretical Foundation

The paper argues that general-purpose language models have advanced in knowledge, reasoning, mathematics, and coding, yet many valuable tasks remain difficult because the work unfolds over a long horizon. The limiting capability is not reasoning in isolation, but the ability to turn reasoning into sustained work inside a changing environment.

The authors define working capability as: sustained, verifiable progress toward a real-world objective, requiring the model to:

• Find and interpret evidence
• Operate on heterogeneous files
• Execute and debug code
• Maintain and revise a coherent plan over many steps
• Recover from failed actions without discarding valid progress
• Deliver artifacts that another person can inspect or continue using

The research is organized around six core principles:

1. Completed Work Is the Unit of Agentic Intelligence – Formalized via a task contract
2. Executable Environments Are a Scaling Surface – Expanding file, search, and code worlds
3. Agentic Coordination Is a Scaling Surface – Training delegation, integration, and replanning
4. A Common Harness Connects Both Dimensions – AgentOS provides persistent runtime
5. Training Exploits Both Scaling Dimensions – SFT and RL over environment and coordination trajectories
6. Evaluation Mirrors the Scaling Design – ReAct (lower scaffold) and Agent Team (system-level lift)

The task contract is formally defined as:

$$\mathcal {E} = (\mathcal {W}, W _ {0}, q, \mathcal {A}, \mathcal {T}, \Omega , \boldsymbol {B}, D, V _ {D}).\tag{1}$$

Where $W$ is the workspace-state space, $W _ { 0 }$ is the initial workspace, $q$ is the objective, $\mathcal {A}$ is the action set, $\mathcal {T}$ is the transition operator, $\Omega$ is the observation interface, $\boldsymbol {B}$ is the resource-budget vector, $D$ is the delivery contract, and $V _ {D}$ is the task-level verifier.

Methodology

Environment Scaling

Apodex 1.1 constructs three complementary environment families, each with distinct assurance boundaries:

The file-world registry spans 33 domains, 318 occupations, and 1,208 deliverable clusters. Difficulty is calibrated using acquisition-pressure coordinates, such as:

$$\rho_ {\mathrm{acq}} (\mathcal {E}) = \frac {N _ {\mathrm{cand}} (\mathcal {E}) + N _ {\mathrm{hop}} (\mathcal {E})}{B _ {\mathrm{tool}} (\mathcal {E})}\tag{6}$$

Where $N _ {\mathrm{cand}}$ is the number of plausible candidates requiring inspection, $N _ {\mathrm{hop}}$ is the number of load-bearing evidence transitions, and $B _ {\mathrm{tool}}$ is the tool-call budget.

Agentic Coordination Scaling: Agent Team

The Agent Team architecture features:

• Explicit Task Board: Decomposition is written to a persistent board with bounded objectives, dependencies, resolution states, and assigned agents
• Asynchronous Human Intervention: User messages during execution are accepted and integrated into the live task board
• Asymmetric Verification: Verification tasks are deliberately narrower than generation—checking specific claims, evidence, and constraints rather than re-solving the entire problem
• Adaptive Max Team Effort: Additional investigation is allocated only to weak, contested, or load-bearing claims
• Evidence-Grounded Synthesis: A dedicated synthesis stage constructs a claim–evidence graph before drafting the final deliverable

AgentOS: Execution Substrate

The workspace state is formalized as:

$$W _ {t} = (F _ {t}, Q _ {t}, C _ {t}, I _ {t}, G _ {t}, K _ {t})\tag{3}$$

Where $F _ {t}$ is file state, $Q _ {t}$ is retrieved evidence, $C _ {t}$ is executable state, $I _ {t}$ is the artifact index, $G _ {t}$ is the dependency graph, and $K _ {t}$ is runtime control state.

Key mechanisms include:

• Stable three-region namespace: /inputs (read-only), /workspace (agents' private work), /outputs (final deliverables)
• Tiered compaction for context pressure management
• Soft/hard budget enforcement with deadline-clamped waits and bounded report recovery
• Controlled artifact delivery via a single-publisher lease and exact manifest with baseline reconciliation

Training

• SFT Stage: Unified mixture spanning general reasoning, tool use, search, file interaction, coding, math, science, financial reasoning, and multi-agent coordination, normalized into a common stateful interaction schema
• Reinforcement Learning with PIVOT-RL: Uses hindsight-guided trajectory localization to identify consequential decision points ("pivots"), preserve useful prefixes, and construct localized continuation tasks with corrective hints
• Asynchronous optimization over heterogeneous agentic workloads

Empirical Validation / Results

Main Results on Professional Work and Finance (Table 3)

Scientific Research and General Reasoning (Table 4)

Mathematical Reasoning (Table 6)

Coding and Software Engineering (Table 7)

Internal Evaluations

FrontierSearchBench (41 verifiable deep-search tasks) — Apodex 1.1 with Agent Team achieves the best average score of 69.1, with 87.8% positive tasks and only 2.4% negative.

FrontierResearchBench (97 executable scientific workflows) — Apodex 1.1 w/ Agent Team achieves a 12.4% Pass Rate, behind GPT-5.6-Sol with Codex (20.6%) and Grok-4.6 with Claude Code (20.6%).

HDS6 Process Analysis

The HDS6 framework evaluates six process capabilities (long-horizon state coherence, evidence fidelity, hypothesis management, boundary/failure reasoning, tool/execution-state management, self-correction under verification) across 24 rubric items, with an integrity gate. Results show the largest single-item deltas in Initial Decomposition (+1.3) and Final Verification (+0.8).

Theoretical and Practical Implications

Theoretical Contributions

1. Working capability as a model-level objective: The paper argues that agentic intelligence should be measured by completed work, not isolated responses, and formalizes this via a task contract ($\mathcal{E}$) that separates initial user input, normalized objective, delivery contract, and verifier.

2. Two scaling dimensions beyond parameters: The paper demonstrates that expanding executable environments and coordination behaviors are consequential scaling surfaces, independent of model parameter count.

3. Training as integration: The paper shows that SFT and RL can convert environment and coordination trajectories into a unified policy, rather than treating tool use and reasoning as disconnected specialist modes.

Practical Implications

1. Model-scale efficiency: The 35B Mini model reaching frontier-band capability suggests that environment design and coordination training may be more effective levers than model size alone for complex work.

2. Execution substrate: AgentOS provides a concrete blueprint for runtime systems that support long-horizon work, including controlled delivery, intervention handling, and context management.

3. Benchmark design: The paper introduces FrontierSearchBench and FrontierResearchBench, which use executable verifiers and cross-artifact consistency checks rather than single-answer grading.

Conclusion

Apodex 1.1 advances a model-level view of agentic intelligence for complex work: capability should be measured by whether a model can sustain progress through a changing task, use tools and evidence effectively, recover from failure, and deliver a verifiable result. The system develops this capability along two complementary scaling dimensions—Environment Scaling and Agentic Coordination Scaling—bound together by a common execution harness (AgentOS) and unified training.

The evaluation positions Apodex 1.1 as a common model and execution stack whose capabilities compose across scientific research, professional work, finance, mathematics, search, and coding. The 35B Mini model demonstrates that these gains are not confined to the largest model.

The longer-term goal is a Heavy-Duty Solver: a system capable of taking responsibility for increasingly ambitious, long-running, and verifiable work. Future directions include scaling environment coverage and fidelity, strengthening learned coordination over longer horizons, improving training and credit assignment over hierarchical traces, and closing the loop between real failures, task construction, training, and evaluation.