Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CORPGEN: Simulating Corporate Environments with Autonomous Digital Employees in Multi-Horizon Task Environments

Abubakarr Jaye, Nigel Boachie Kumankumah, Chidera Biringa, Anjel Shaileshbhai Patel, Sulaiman Vesal, Dayquan Julienne, Charlotte Siska, Manuel Raúl Meléndez Luján, Anthony Twum-Barimah, Mauricio Velazco, Tianwei Chen

Abstract

Long-horizon reasoning is a key challenge for autonomous agents, yet existing benchmarks evaluate agents on single tasks in isolation. Real organizational work requires managing many concurrent long-horizon tasks with interleaving, dependencies, and reprioritization. We introduce Multi-Horizon Task Environments (MHTEs): a distinct problem class requiring coherent execution across dozens of interleaved tasks (45+, 500-1500+ steps) within persistent execution contexts spanning hours. We identify four failure modes that cause baseline CUAs to degrade from 16.7% to 8.7% completion as load scales 25% to 100%, a pattern consistent across three independent implementations. These failure modes are context saturation (O(N) vs O(1) growth), memory interference, dependency complexity (DAGs vs. chains), and reprioritization overhead. We present CorpGen, an architecture-agnostic framework addressing these failures via hierarchical planning for multi-horizon goal alignment, sub-agent isolation preventing cross-task contamination, tiered memory (working, structured, semantic), and adaptive summarization. CorpGen simulates corporate environments through digital employees with persistent identities and realistic schedules. Across three CUA backends (UFO2, OpenAI CUA, hierarchical) on OSWorld Office, CorpGen achieves up to 3.5x improvement over baselines (15.2% vs 4.3%) with stable performance under increasing load, confirming that gains stem from architectural mechanisms rather than specific CUA implementations. Ablation studies show experiential learning provides the largest gains.

CORPGEN: Simulating Corporate Environments with Autonomous Digital Employees in Multi-Horizon Task Environments

Abstract

Long-horizon reasoning is a key challenge for autonomous agents, yet existing benchmarks evaluate agents on single tasks in isolation. Real organizational work requires managing many concurrent long-horizon tasks with interleaving, dependencies, and reprioritization. We introduce Multi-Horizon Task Environments (MHTEs): a distinct problem class requiring coherent execution across dozens of interleaved tasks (45+, 500-1500+ steps) within persistent execution contexts spanning hours. We identify four failure modes that cause baseline CUAs to degrade from 16.7% to 8.7% completion as load scales 25% to 100%, a pattern consistent across three independent implementations. These failure modes are context saturation (O(N) vs O(1) growth), memory interference, dependency complexity (DAGs vs. chains), and reprioritization overhead. We present CorpGen, an architecture-agnostic framework addressing these failures via hierarchical planning for multi-horizon goal alignment, sub-agent isolation preventing cross-task contamination, tiered memory (working, structured, semantic), and adaptive summarization. CorpGen simulates corporate environments through digital employees with persistent identities and realistic schedules. Across three CUA backends (UFO2, OpenAI CUA, hierarchical) on OSWorld Office, CorpGen achieves up to 3.5x improvement over baselines (15.2% vs 4.3%) with stable performance under increasing load, confirming that gains stem from architectural mechanisms rather than specific CUA implementations. Ablation studies show experiential learning provides the largest gains.
Paper Structure (27 sections, 3 figures, 4 tables, 1 algorithm)

This paper contains 27 sections, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Setting
  5. Why Single-Task Long-Horizon Agents Fail in MHTEs
  6. Required Capabilities: Multi-Objective Multi-Horizon Agents (MOMA)
  7. CorpGen Framework
  8. Hierarchical Planning for Long-Term Consistency
  9. Sub-Agents as Tools for Context Isolation
  10. Tiered Memory Architecture
  11. Summarization and Context-Management Modules
  12. Cognitive Tools for Structured Operations
  13. Experiential Learning
  14. Emergent Collaboration
  15. Experimental Results
  16. ...and 12 more sections

Figures (3)

  • Figure 1: A Digital Employee's Workday. The agent begins with Day Init, loading memory and generating a daily plan. It then runs repeated Execution Cycles, where it retrieves context, reasons and acts, and persists results while completing many interleaved tasks. At Day End, the agent reflects on its actions and consolidates experience into long-term memory, enabling coherent operation over hours despite context resets.
  • Figure 2: CorpGen Architecture. Four core mechanisms address long-horizon challenges: Hierarchical Planning (§\ref{['subsec:hierarchical_planning']}) decomposes strategic objectives (monthly) into tactical plans (daily) and operational actions (per-cycle) to prevent goal drift. The Execution Engine runs a ReAct loop (Algorithm \ref{['alg:agent_execution']}, lines 11--15) invoking Cognitive Tools (§\ref{['subsec:cognitive_tools']}) within host context and Sub-Agents as Tools (§\ref{['subsec:subagents']}) in isolated context scopes. Context Management (§\ref{['subsec:context_management']}) prevents overflow via adaptive summarization. Tiered Memory (§\ref{['subsec:memory_arch']}) prevents information loss across working memory (intra-cycle), structured LTM (plans, summaries), and semantic memory (Mem0 embeddings). Experiential Learning (§\ref{['subsec:experiential_learning']}) captures successful execution patterns via UFO2 trajectory learning, with feedback routed directly to UFO2's app agent. External interfaces include enterprise applications via UFO2 GUI and communication channels enabling Emergent Collaboration (§\ref{['subsec:collaboration']}).
  • Figure 3: Experiential learning cycle. UFO2's trajectory learning captures successful execution patterns and distills them into canonical demonstrations indexed for retrieval. Feedback from successful executions is routed directly to the execution agent, biasing future action selection toward validated patterns.