Skip to content

Can LLM Agents Sustain Long-Horizon Organizational Dynamics?

🕒 Published (v1): 2026-05-31 12:28 UTC · Source: Arxiv · link

Why this paper was selected

Tests whether LLM agents sustain coherent behavior in structured organizational hierarchies

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TaskWeave is a hierarchical agentic framework that frames long-horizon organizational simulation as a memory-centered coordination problem, maintaining coherent role delegation, plan propagation, and execution traceability across year-long multi-agent simulations. It introduces a Formulate–Partition–Diagnose–Align (FPDA) planning cycle and dependency-aware trace memory to keep planning and execution states coherent across temporal layers. Experiments on a simulated SaaS company show it outperforms existing multi-agent baselines on organizational coherence and grounded artifact utility.

Problem

Existing LLM multi-agent frameworks (AutoGen, MetaGPT, CAMEL, etc.) handle short-horizon task collaboration or open-ended social interaction but cannot sustain the three coupled coherence dimensions of organizational simulation: (1) role-structured collaboration through explicit delegation paths, (2) hierarchical intent propagation from yearly objectives to weekly tasks, and (3) dependency-grounded execution where actions are anchored in prior outputs and external evidence. Without persistent planning-state memory, agents produce locally plausible but globally inconsistent behavior — role assignments drift, task dependencies go unresolved, and earlier plans are not inherited by later cycles.

Method

TaskWeave has three components operating over a shared simulation memory:

Organizational Prior Instantiation compiles domain metadata \(B\) into a population \(A = \{a_i\}_{i=1}^N\) where each agent \(a_i = (\phi_i, \rho_i)\) holds a personal profile \(\phi_i\) and a role descriptor \(\rho_i\). Agents are arranged in \(T\) tiers and connected by a delegation graph \(H = (A, E_H)\) that constrains valid task routing.

Hierarchical Intent Propagation (FPDA) maintains planning states \(S_{l,i}^p = (\pi_{l,i}^p, T_{l,i}^p, \sigma_{l,i}^p)\) indexed by temporal level \(l\) (year/quarter/month/week) and planning unit \(i\). The FPDA transition \(S_{l+1,j}^i = \mathcal{F}_{\text{FPDA}}(S_{l,i}^p, A^{[t]}, H)\) sequentially: (F) formulates a local plan \(\pi' = \mathcal{F}_{\text{form}}(\pi, \sigma, B)\); (P) partitions it into a task bundle \(T' = \mathcal{F}_{\text{part}}(\pi', A^{[t]}, H)\); (D) derives a diagnostic summary \(\sigma' = \mathcal{F}_{\text{diag}}(R, T')\) from execution results; (A) aligns the plan \(\tilde{\pi}' = \mathcal{F}_{\text{align}}(\pi', \sigma')\). Cross-unit sibling states synchronize via an alignment operator \(\mathcal{G}_{\text{align}}\); feedback aggregates upward through \(\mathcal{U}_{\text{up}}(\{\sigma_{l+1,j}^i\}_{j \in C(i)})\).

Dependency-Aware Operational Execution decomposes composite tasks \(\tau_{l,i,m} = (\eta_m, \delta_m, \xi_m)\) into atomic subtasks \(\hat{\tau}_{m,k} = (\hat{\eta}_{m,k}, \Psi_{m,k})\), where \(\Psi_{m,k}\) is a structured dependency query. Each query resolves a context \(\Gamma_{m,k} = \Gamma_{m,k}^{\text{int}} \cup \Gamma_{m,k}^{\text{ext}}\) from the global result pool \(R\) and tool space \(\Omega\). A role-compatible agent \(a^*_{m,k} = \arg\max_{a : (a_{\text{src}},a) \in E_H} s_{\text{match}}(a, \hat{\eta}_{m,k})\) is selected by semantic compatibility. Results are written back as \(R \leftarrow R \cup \{(\eta_m, \delta_m, \xi_m)\}\), forming traceable process memory for subsequent FPDA cycles.

Key Contributions

  • Formal definition of long-horizon organizational simulation as a memory-centered coordination problem spanning role consistency, intent propagation, dependency satisfaction, and traceable execution.
  • TaskWeave: role-structured agent population + FPDA planning-state memory + dependency-aware trace memory, all over a shared delegation graph.
  • Organizational Sensitive Span Detection (OSSD) as a downstream evaluation task measuring whether generated artifacts contain structured, context-dependent enterprise-sensitive content.
  • Year-long SaaS simulation testbed with 15 agents, 4-level temporal hierarchy, 6 LLM backbone ablations, and comparison against 6 multi-agent baselines.

Results

  • Plan propagation (monthly): Finalized completion reaches 92.66% (Gemini), 91.39% (ChatGPT), 82.01% (Deepseek), 83.79% (Moonshot); LLaMA3 and GLM fail to produce sufficiently structured plans for checklist evaluation.
  • Plan propagation (weekly): Finalized completion reaches 98.58% (Gemini), 95.32% (ChatGPT), 92.55% (Deepseek), 91.72% (Moonshot), 68.21% (LLaMA3).
  • Organizational coherence (KL divergence from reference role prior): Gemini 0.144, Deepseek 0.130, Moonshot 0.208, ChatGPT 0.274, LLaMA3 0.442, GLM 0.502; stronger backbones achieve broader role coverage.
  • OSSD utility (single-stage baseline comparison): TaskWeave achieves diversity score 31 and span length 643 using only 5.6 API calls and 30k/4k prompt/completion tokens, versus Magentic-One (31.0 API calls, 465k/19k tokens, diversity 83) and G-Designer (40.6 API calls, 16k/25k tokens, diversity 19).
  • Responsiveness: 534 tool calls across the year-long run; external events (economic, technology) measurably shift workload distribution toward corresponding departments.
  • Transferability: Framework successfully instantiated on financial company, manufacturing plant, and government agency settings (detailed results truncated in paper text).

Limitations

  • Evaluation is confined to a single simulated SaaS company (Company A, 15 agents) as the primary testbed; transferability results are qualitative.
  • Weaker LLM backbones (LLaMA3, GLM) fail to sustain stable plan propagation, making the framework backbone-dependent rather than universally applicable.
  • The reference role prior used for KL-divergence evaluation is expert-designed and domain-specific, limiting its use as a universal organizational coherence metric.
  • OSSD comparison with baselines uses a shared single-stage execution setting that does not exercise the full year–quarter–month–week hierarchy for baseline methods.
  • No evaluation of computational cost or latency scaling as simulation horizon or agent count grows.

Relevance to Harnesses / Meta-Harnesses

TaskWeave is directly a meta-harness: it orchestrates populations of LLM agents through an explicit, parameterized control loop (FPDA) rather than relying on emergent coordination. The FPDA cycle — plan, partition, diagnose, align — is structurally analogous to the outer planning shell in research harnesses that decompose a high-level goal into subtasks, monitor execution, and revise the plan from feedback. The dependency-aware trace memory and global result pool demonstrate a concrete pattern for how a harness can maintain grounded, provenance-tagged state across long multi-agent runs, directly informing harness designs that must remain coherent over dozens or hundreds of agent calls. The paper's finding that intent preservation, not task generation, is the bottleneck in long-horizon multi-agent systems is a key design principle for anyone building orchestration harnesses that span extended horizons.