Technical Summary: TaskWeave for Long-Horizon Organizational Dynamics

1. Problem Statement

While Large Language Model (LLM) agents have advanced in simulating social and economic interactions, they struggle to sustain coherent behavior in structured organizations over long time horizons. Existing frameworks often fail to maintain organization-level state across hierarchy, time, and execution dependencies. Specifically, agents in current systems tend to generate locally plausible outputs while gradually losing consistency in:

• Role assignment: Agents may drift from their defined responsibilities.
• Task dependencies: Later decisions may contradict earlier plans or ignore prerequisite artifacts.
• Plan inheritance: High-level goals are not effectively propagated or preserved as they are decomposed.
• Execution history: Actions are often treated as isolated responses rather than grounded in prior process traces.

The paper defines this as the long-horizon organizational coherence problem, where a simulator must preserve the evolving organizational state behind locally plausible actions.

2. Methodology: TaskWeave

The authors propose TaskWeave, a hierarchical agentic framework designed to sustain long-horizon organizational dynamics through a memory-centered coordination approach. The framework models an organization as a role-structured agent population constrained by an explicit delegation topology.

Core Components

TaskWeave consists of three primary mechanisms:

A. Organizational Prior Instantiation

Instead of treating agents as interchangeable workers, TaskWeave compiles organizational metadata (domain, scale, departments, role taxonomy) into a structured agent population.

• Agent Representation: Each agent $a_i$ is defined by personal attributes ($\phi_i$, e.g., background) and functional role descriptions ($\rho_i$, e.g., title, responsibilities, privileges).
• Delegation Topology: Agents are partitioned into tiers (e.g., CEO, Managers, Workers) and connected via a delegation graph $H$. This graph defines valid coordination paths and constraints on how intent propagates and which agents can receive tasks.

B. Hierarchical Intent Propagation (FPDA Cycle)

To maintain planning states across temporal levels (e.g., Year $\to$ Quarter $\to$ Month $\to$ Week), TaskWeave employs a Formulate–Partition–Diagnose–Align (FPDA) cycle.

• Planning State: Defined as $S_{l,i} = (\pi, T, \sigma)$, where $\pi$ is the plan, $T$ is the task bundle, and $\sigma$ is a diagnostic summary from downstream execution.
• The Cycle:
  1. Formulate: Derives a local plan from upstream intent and diagnostic context.
  2. Partition: Converts the plan into a task bundle delegable through the topology $H$.
  3. Diagnose: Summarizes execution results, deviations, and blockers from the result pool.
  4. Align: Revises the local plan using diagnostic feedback to ensure consistency with higher-level goals.
• Feedback Loops: The framework supports cross-unit synchronization (e.g., deadline adjustments between siblings) and upward feedback aggregation, ensuring that execution evidence updates future planning states.

C. Dependency-Aware Operational Execution

Execution is not stateless generation; it is conditioned on role constraints, resolved dependencies, and traceable memory.

• Task Decomposition: Composite tasks are broken into atomic subtasks with explicit dependency queries ($\Psi$).
• Context Resolution: Before execution, the system resolves dependencies by retrieving internal results ($\Gamma_{int}$) from the global result pool and external evidence ($\Gamma_{ext}$) via tools.
• Agent Selection: Subtasks are assigned to agents based on role compatibility and authority within the delegation graph.
• Traceable Memory: Execution results are written back to a global result pool as traceable process memory, preserving provenance (responsible agent, dependencies, tool interactions). This allows future tasks to retrieve and reuse organizational context.

3. Experimental Setup

The authors evaluated TaskWeave in a year-long simulation of a SaaS company (Company A) involving 15 agents across 3 hierarchical tiers and 3 departments.

• Backbones: Evaluated across six LLMs (GPT-4o-mini, Gemini-2.0-Flash, Deepseek-v3, Moonshot-v1-8K, LLaMA3.1-70B, GLM-4-Flash).
• Baselines: Compared against AutoGen, MetaGPT, CAMEL, MoA, Magentic-One, and G-Designer.
• Evaluation Metrics: Organizational coherence (role consistency), plan propagation (intent preservation), grounded execution (artifact reusability), enterprise NLP utility (sensitive span detection), and responsiveness to external events.

4. Key Results

The experiments yielded six primary observations:

1. Role Activation: While organizational structure guides delegation, the capacity of the LLM backbone determines whether roles are actually activated. Stronger models (Gemini, Deepseek) achieved broader role coverage and lower divergence from ideal role distributions compared to weaker models (GLM).
2. Intent Preservation: Long-horizon simulation is constrained by intent preservation rather than just plan generation. Stronger backbones maintained high "finalized completion" rates across monthly and weekly layers, whereas weaker models failed to produce structured plans or lost intent over time. The gap between "timely" and "finalized" completion highlights the system's ability to handle organizational spillover.
3. Grounded Execution: TaskWeave produces workflows where intermediate artifacts are reusable process memory. Generated outputs (plans, documents, code) are linked via dependency graphs, enabling later tasks to retrieve and build upon prior context.
4. Enterprise NLP Utility: In terms of generating enterprise-sensitive content (e.g., compliance data, financial spans), TaskWeave achieved a superior efficiency–richness trade-off. It produced more sensitive spans with fewer API calls and lower token costs compared to general-purpose frameworks like AutoGen and MetaGPT.
5. Responsiveness: The system supports bidirectional environment interaction. Injected external incidents (policy, economic, technology) induced interpretable shifts in strategic workloads, and agents successfully executed tool calls (SQL, email, social media) to act on the environment.
6. Transferability: The framework successfully transferred to diverse domains (Financial, Manufacturing, Government) with varying scales and hierarchy depths, maintaining stable completion rates as organizational complexity increased.

5. Significance and Claims

The paper claims that structured simulation memory is a key mechanism for building reliable LLM-based organizational simulators.

• Contribution to Simulation: TaskWeave demonstrates that explicit hierarchy, dependency tracking, and traceable execution are sufficient for LLM agents to produce plausible coordination patterns and grounded process traces over long horizons.
• Practical Utility: The framework serves as a controllable environment for studying multi-agent dynamics without relying on costly or private real-world experiments. It generates high-quality, process-grounded artifacts useful for downstream enterprise NLP analysis.
• Limitations: The authors acknowledge that the simulation remains a simplified abstraction of real organizations. Future work is needed to incorporate richer institutional rules, market signals, and human-in-the-loop calibration. The framework is positioned as a controllable simulator for studying dynamics under structural constraints, not as a literal digital twin of a specific real-world company.

In summary, TaskWeave addresses the gap between short-horizon task solving and long-horizon organizational simulation by introducing a memory-centered architecture that couples hierarchical planning with dependency-aware execution.