AgentAsk: Clarification in MAS

• AgentAsk is a lightweight clarification module for LLM-based multi-agent systems that detects and mitigates inter-agent message failures.
• It employs a three-stage pipeline: distilling failure traces, supervised fine-tuning of a clarification policy, and online E-GRPO optimization under strict resource constraints.
• Empirical evaluations show AgentAsk improves system accuracy and efficiency by effectively addressing prevalent edge-level errors like Data Gap and Signal Corruption.

AgentAsk is a lightweight, plug-and-play clarification module for large-language-model-based multi-agent systems (MAS) that treats every inter-agent message as a potential failure point and inserts minimally necessary questions to arrest error propagation (Lin et al., 8 Oct 2025). It is motivated by the claim that MAS frequently underperform single-agent baselines because minor inaccuracies at one message handoff propagate across the entire chain as edge-level error cascades. Rather than replacing an existing orchestration stack, AgentAsk is designed as an architecture-agnostic link-local intervention layer that distills edge-level judgments from curated failure traces, supervises a policy for deciding when, what, whom, and how to ask, and then optimizes that policy online with E-GRPO under explicit accuracy, latency, and cost constraints (Lin et al., 8 Oct 2025).

1. Problem formulation and motivation

The central problem addressed by AgentAsk is reliability at the agent-to-agent handoff. In the formulation given for the system, the dominant failure mode in LLM-based MAS is not only incorrect final reasoning, but the propagation of small local defects in messages exchanged between agents. Analysis of 824 MAS execution logs showed that failures mostly originate at the information handoff between agents, and that even sophisticated multi-agent orchestration, role design, and self-correction often fail to catch or contain these errors efficiently (Lin et al., 8 Oct 2025).

This framing shifts the intervention point from global evaluation to the communication edge itself. Each handoff is treated as a potential site of ambiguity, omission, corruption, or role mismatch. AgentAsk therefore operates locally: it inspects the edge context, determines whether a clarification is necessary, selects an addressee, and issues a minimal question intended to prevent downstream compounding. The design goal is not maximal questioning, but minimally necessary questioning under resource budgets.

2. Edge-level error taxonomy

A principal contribution of AgentAsk is a four-type taxonomy of edge-level errors at inter-agent message handoffs (Lin et al., 8 Oct 2025). The taxonomy is intended to support both diagnosis and intervention, and it also structures the supervised and reinforcement-learning stages of the method.

Error type	Description	Prevalence
Data Gap (DG)	Required information is missing	29.1%
Referential Drift (RD)	Symbols, names, or references become ambiguous or inconsistent	27.3%
Signal Corruption (SC)	Intermediate values, formats, or units are corrupted or distorted	36.8%
Capability Gap (CG)	The designated role or agent lacks the skills or tools for the subtask	6.8%

Signal Corruption is reported as the most prevalent category at 36.8%, followed by Data Gap at 29.1%, Referential Drift at 27.3%, and Capability Gap at 6.8%. The taxonomy is explicitly edge-local: Data Gap captures missing boundary cases or implicit assumptions; Referential Drift captures unstable reference chains such as pronouns or reused indices; Signal Corruption captures distorted intermediate values, formats, or units; Capability Gap captures role assignment failures in which the designated agent lacks the necessary competence or tools.

3. Three-stage pipeline

AgentAsk follows a three-stage pipeline (Lin et al., 8 Oct 2025). In the first stage, knowledge is distilled from failure traces. The objective is to build a high-quality dataset of edge-level intervention examples by tracing where MAS fail and using a powerful evaluator as a “teacher” to generate minimal clarifications. The resulting supervised fine-tuning corpus is written as

$\widetilde{\mathcal{D}} = \{(x_i, y_i)\}_{i=1}^N,$

where each $x_i$ is the edge context and

$y_i = (t_i, v_i, q_i)$

contains the error type, addressee, and minimal clarification question.

In the second stage, a lightweight controller is trained by supervised fine-tuning to imitate the teacher and determine when, what, whom, and how to ask. The controller is constrained by a schema and budget, reflecting the design requirement that clarifications be brief, targeted, and cost-aware.

In the third stage, the policy is optimized online with E-GRPO. This stage adapts the ask policy to maximize task utility while respecting latency and token-cost constraints. The optimization objective explicitly balances edge-local gains with final task success, which makes clarification a learned intervention policy rather than a fixed heuristic.

4. Policy structure and optimization

The policy is factorized into an ask gate, an addressee distribution, and a question generator (Lin et al., 8 Oct 2025):

$$\pi_\theta(a_t \mid x_t) = \pi_\theta(z_t \mid x_t)\left[\pi_\theta(\tilde{v}_t \mid x_t)\pi_\theta(q_t \mid x_t)\right]^{z_t},$$

where $z_t \in \{0,1\}$ is the ask gate, $\tilde{v}_t$ is the addressee, and $q_t$ is the clarification question. This factorization makes the intervention explicit: the system first decides whether to ask, and only then decides whom to ask and how to phrase the clarification.

The supervised phase uses the loss

$$\mathcal{L}_{\mathrm{SFT}} = \mathcal{L}_{\mathrm{type}} + \lambda_{\mathrm{ask}}\mathcal{L}_{\mathrm{ask}},$$

combining cross-entropy for error-type classification with conditioned text loss for addressee and question generation. The controller output is schema-constrained to an ask/no-ask decision, an error type, an addressee, and a minimal clarification.

The reinforcement stage defines an effectiveness reward that is positive only when asking removes downstream uncertainty:

$$r_t^{\mathrm{eff}} = \begin{cases} +1, & z_t=1 \wedge n_{t+1}=0 \ -1, & z_t=1 \wedge n_{t+1}=1 \ 0, & z_t=0 \end{cases}$$

A parsimony term penalizes gratuitous or excessive asking,

$$r_t^{\mathrm{par}} = -\lambda_{\mathrm{sw}}\max(c_t-1,0),$$

and a format bonus rewards schema compliance and briefness,

$x_i$0

These are combined as

$x_i$1

with terminal task success reward

$x_i$2

The constrained optimization perspective is stated as

$x_i$3

The E-GRPO objective then combines local and global advantages with clipping and a KL trust region to the supervised reference policy:

$x_i$4

This optimization directly encodes the system’s intended trade-off: maximal error arrest per clarification subject to bounded overhead.

5. Empirical evaluation

AgentAsk is evaluated across math, reasoning, and coding benchmarks, specifically GSM8K, MATH, MMLU, HumanEval, and MBPP, and is tested on four public multi-agent frameworks—GPTSwarm, AFlow, MaAS, and MasRouter—alongside single-agent baselines such as IO and CoT (Lin et al., 8 Oct 2025). Across these settings, AgentAsk improves accuracy in approximately 85% of framework-by-dataset cases, reported as 17 out of 20, with a mean increase of about 0.5 percentage-points over the original MAS.

The module is also reported to approach the performance of a GPT-5-powered clarifier while operating at less than one-eighth of the cost. Overhead remains small: the paper states that latency and extra cost are both less than 5% over the original MAS, and presents an efficiency frontier in which AgentAsk is near Pareto-optimal relative to heavier evaluators.

A sample result on MasRouter for GSM8K illustrates the trade-off:

Variant	Accuracy	Latency / Extra cost
Origin	93.26%	100% / 0%
+AgentAsk	94.72%	103% / 4.2%
+GPT-5	95.34%	134% / 38.0%

The gains are reported as particularly strong for error types most amenable to one-turn clarification, especially Data Gap and Signal Corruption. The paper also notes occasional no-gain or slight-regression cases when failures are dominated by Referential Drift or Capability Gap, which are less one-shot fixable. This distinction is consistent with the taxonomy: some handoff failures are locally correctable by a single brief clarification, whereas others reflect deeper representational or assignment problems.

6. Position within the broader literature on asking and agent reliability

AgentAsk addresses a specific form of asking: inter-agent clarification at communication edges. This focus differs from uncertainty-aware multi-agent coding systems such as the scaffold in "Ask or Assume? Uncertainty-Aware Clarification-Seeking in Coding Agents," where an Intent Agent detects underspecification and a Main Agent pauses execution to query the user; in that setting, the decoupled multi-agent system reaches a 69.40% task resolve rate versus 61.20% for a standard single-agent setup (Edwards et al., 27 Mar 2026). It also differs from benchmarks such as HiL-Bench, which measure selective escalation to humans under missing information, ambiguity, and contradiction using Ask-F1 rather than link-local handoff repair (Elfeki et al., 10 Apr 2026).

Within multi-agent architecture research, AgentAsk can be read alongside evidence that orchestration choices alone do not resolve systemic failure. The enterprise benchmark AgentArch reports significant model-specific architectural preferences and low absolute reliability, with maximum success of only 35.3% on the more complex task and 70.8% on the simpler task across 18 configurations (Bogavelli et al., 13 Sep 2025). Likewise, M-ASK addresses stability by explicitly decoupling search planning from knowledge management and assigning turn-level rewards to both search decisions and knowledge updates (Chen et al., 8 Jan 2026). This suggests that asking, decomposition, and reward shaping are complementary responses to different bottlenecks in agent systems.

In that broader landscape, AgentAsk’s distinctive claim is that many multi-agent failures are edge-local and can be mitigated without redesigning the full system. Its taxonomy of Data Gap, Referential Drift, Signal Corruption, and Capability Gap, together with a plug-in controller trained by distillation, supervision, and E-GRPO, makes clarification an explicit systems primitive for MAS reliability rather than an ad hoc conversational behavior (Lin et al., 8 Oct 2025).