Multi-agent systems powered by large language models (LLMs) can produce incorrect answers, but not all failures are alike. According to a new paper on arXiv titled "Trustworthy Multi-Agent Systems: Mitigating Semantic Drift with the Argent Signaling Protocol," some answers are grounded in the right material yet incomplete, while others are entirely ungrounded and should be halted. Current retry strategies treat both cases identically—simply trying again—leaving human supervisors unable to determine whether a retry was warranted or whether the system should have stopped. To address this, the researchers introduce the Argent Signaling Protocol (ASP), a compact machine-readable header that accompanies every AI-generated response with structured quality signals.
The Problem of Semantic Drift
When multiple LLM agents collaborate, information can degrade as it passes between agents—a phenomenon known as semantic drift. The paper notes that current systems have no mechanism to communicate the reliability of an output, leading to blind retries that propagate errors. ASP aims to solve this by embedding four signals into each response:
- @C (Certainty): How confident the model is in its answer.
- @G (Grounding): Whether the response is grounded in source material.
- @S (Stochasticity): Variation in outputs due to model randomness.
- Assumption Index: Classifies the evidentiary basis of each claim.
These signals enable a controller to distinguish repairable failures (grounded but incomplete) from containment failures (ungrounded) and route each case differently—either retry with adjusted parameters or halt completely.
Experimental Evaluation
The researchers evaluated ASP in two modes: standalone and multi-agent. The benchmark used a 27-question document-grounded QA task over the Array BioPharma/Ono license agreement. Three local GGUF models were tested.
| Model | Baseline Pass Rate | ASP Pass Rate | Improvement |
|---|---|---|---|
| Qwen (0.8B) | 11.1% | 33.3% | +22.2 pp |
| Dobby (8B) | 33.3% | 44.4% | +11.1 pp |
| SmolLM3 (3B) | Variable per question | Alternated repair/containment | N/A |
| Aggregate | 12/81 | 21/81 | +11.1% |
On Qwen (0.8B), ASP also raised mean term coverage from 36.7% to 65.4%. On Dobby (8B), ASP produced 4 fail-to-pass recoveries. SmolLM3 (3B) showed ASP's ability to alternate between repair and containment per question.
In multi-agent mode, an ASP sidecar sat between a retrieval agent and a downstream decision agent. The sidecar blocked 100% of ungrounded upstream outputs from reaching the downstream agent—24 out of 27 upstream outputs were blocked, and 0 ungrounded propagations occurred.
Implications for Enterprise AI
For enterprise technology leaders, the ability to distinguish failure types in multi-agent LLM systems is critical for building trustworthy AI pipelines. ASP provides a lightweight, structured mechanism to audit AI outputs and decide when to retry versus when to escalate to a human. The protocol could be integrated into existing agent architectures, offering a layer of quality control that currently relies on opaque retry loops. While the evaluation is limited to a single license-agreement benchmark, the results suggest that ASP can significantly reduce the propagation of ungrounded information, a key requirement for regulated industries such as finance, healthcare, and legal document processing.