MiTa: Hierarchical Multi-Agent Framework

• MiTa is a hierarchical, memory-integrated multi-agent framework that uses a manager–member architecture to coordinate global task allocation and ensure memory consistency via episodic summaries.
• It employs a global task allocation mechanism that integrates individual agent proposals and compressed episodic memories to prevent inter-agent conflicts and boost efficiency.
• Experimental benchmarks show that MiTa significantly reduces task steps and improves scalability compared to previous LLM-based systems through robust global reasoning.

MiTa is a hierarchical, memory-integrated multi-agent collaboration framework designed to address inter-agent conflict and memory inconsistency in LLM-driven embodied systems. The architecture organizes agents into a top-down manager–member structure, equipping the manager with two unique modules—Allocation for global task coordination and Summary for episodic memory integration—enabling efficient, consistent cooperation on complex tasks. Experimental benchmarks with multiple agent configurations demonstrate MiTa's superior efficiency and adaptability compared to previous LLM-based multi-agent systems (Zhang et al., 30 Jan 2026).

1. Hierarchical Manager–Member System

MiTa implements a strict manager–member hierarchy. The manager centralizes decision-making, maintaining a global belief state $b_t(s)$ over the joint environment and task objective $T$. Members individually perceive local observations $o_i^t$, encode them (structured or raw RGB-D), and propose candidate actions $m_i^t$ via an LLM-driven negotiation module:

$$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$

Each member preserves its own short-term memory $\mathcal H_i$ and executes the manager-assigned subtask $a_i^t$. The manager extends standard modules with Allocation (for task division and conflict avoidance) and Summary (for compressed, task-triggered memory updates).

2. Global Task Allocation Mechanism

Task allocation is formalized in the MPOMDP framework with joint action space $\mathbb{A}^d = A_1 \times ... \times A_n$. At each step, the manager aggregates member proposals, beliefs, and observations into a cross-agent context:

$$\mathcal X^t = \bigcup_{i=1}^n \{m_i^t, b_i^t, o_i^t\}$$

Allocation is performed by selecting the joint action maximizing a global scoring function instantiated by the LLM:

$$\mathbf a^{t*} = \arg\max_{\mathbf a^t \in \mathbb{A}^d} \mathcal M(\mathcal X^t, \mathcal C, p^t, T)$$

where $T$0 is implemented as an LLM prompt incorporating collaborative summary $T$1 and task progress $T$2. The allocation process explicitly avoids conflicts by operating from a global perspective. Algorithmically, the context is assembled, the LLM ranks feasible joint actions, and the top-scoring plan is dispatched to members.

3. Episodic Memory Integration via Summarization

MiTa addresses long-horizon context retention with an episodic summary module. Both manager and members log each executed joint action $T$3 and belief $T$4. When task progress, quantified via $T$5, changes, the manager triggers a summarization routine:

• Retrieve recent history chunk:

$T$6

• LLM prompt uses the progress delta $T$7 and $T$8 to compress recent events into summary $T$9:

$o_i^t$0

This mechanism ensures the manager's cross-agent context for allocation always incorporates the latest condensed sequence of collaboration, enabling consistent and scalable context propagation without unbounded memory growth.

4. Time Evolution and Consistency Maintenance

The operational loop of MiTa is organized such that, at each time step:

• Members generate and send $o_i^t$1 to the manager.
• The manager, using $o_i^t$2, allocates $o_i^t$3 globally.
• Members act in the environment; the joint outcome ($o_i^t$4) is observed.
• Beliefs are updated:

$o_i^t$5

• If $o_i^t$6, as evaluated against $o_i^t$7, changes, the summary module is invoked and $o_i^t$8 is refreshed. This process ensures that memory fragments propagate forward efficiently and all task allocations remain globally coherent, with summaries available for future reference and planning.

5. Computational and Memory Complexity

The allocation module operates at time complexity $o_i^t$9 per step ($m_i^t$0 agents, $m_i^t$1 candidate joint actions; $m_i^t$2 is pruned in practice via prompt restrictions), with memory $m_i^t$3 for proposals and $m_i^t$4 for episodic summaries. The summary module triggers only on progress changes ($m_i^t$5 number of subtasks), operates in $m_i^t$6 time per summary ($m_i^t$7 = summary history window), and transiently requires $m_i^t$8 memory.

6. Empirical Performance and Robustness

Experimental validation in VirtualHome-Social (C-WAH benchmark, five household tasks) demonstrates MiTa's superiority over CoELA and ProAgent. For three agents (GPT-4o backbone, symbolic mode), MiTa achieves an average step count $m_i^t$9 ($$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$0 efficiency improvement), outperforming CoELA ($$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$1) and ProAgent ($$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$2). Across all task types and agent counts, MiTa delivers the lowest step count, with its hierarchical allocation and episodic memory structures yielding the largest gains in the multi-agent case. Robustness testing shows downgrading member LLMs marginally increases $$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$3 ($$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$4+1), but downgrading the manager LLM increases $$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$5 by $$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$6–$$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$7, confirming high-quality global reasoning is critical (Zhang et al., 30 Jan 2026).

7. Module Ablation and Sensitivity

Ablation studies confirm that removal of the allocation module (leaving only member negotiation) increases steps by $$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$8, leading to uncoordinated, myopic behaviors. Removing the summary module (no episodic integration) increases steps by $$m_i^t = f_{\mathrm{LLM}}(o_i^t, \mathcal H_i, p^t, T)$$9, indicating degraded long-horizon context and increased action conflict. Full MiTa, with both modules, attains maximal efficiency.

---

MiTa represents an overview of negotiation-aware, hierarchical task allocation with episodic, LLM-driven memory summarization, systematically addressing coordination and memory bottlenecks in LLM-based multi-agent systems. Empirical evidence confirms its efficiency and robustness, particularly in complex multi-agent environments requiring long-horizon, globally consistent planning (Zhang et al., 30 Jan 2026).