AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
Jaewon Chu, Jinwoo Seo, Jaewon Cho, Jeehye Na, Yunyang Xiong, Youngdae Kim, Hyunwoo J. Kim
Korea University · KAIST · Meta AI · UNIST
Published September 8, 2026 · on the feed September 10, 2026 · ▲ 82 · arXiv 2609.08572
The abstract
Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a target prompt without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose AgentGrad, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention modifies the behavior of one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks and reduces wall-clock optimization time by 2.5times on average compared to the next-fastest baseline.
The Brief
2:43A short podcast summary of the paper. Written and voiced by AI.
Transcript
AgentGrad. Intervention-guided Prompt Optimization for Multi Agent Systems. By Jaewon Chu and six others, from Korea University, KAIST, and two other institutions. Published on the eighth of September. No code repository is linked to it.
Multi-agent systems built on language models, MAS for short, get strong results by assigning specialised roles to several agents. Their performance, though, depends on how each agent's prompt is written. The leading way to optimize those prompts is the textual gradient: natural-language feedback that guides how a prompt should be changed.
The authors find limits at two stages of existing textual-gradient methods. The first is extraction. Previous approaches pick a target prompt without checking whether changing it would actually fix the failure, and they derive gradients without supervision at the level of the agent, meaning without a view of what that agent's intermediate output should have been. The second is aggregation. Individual gradients are grouped at random and concatenated, which often mixes unrelated failure modes and yields prompts that fail to generalize. Put together, the two gaps mean feedback can be aimed at the wrong agent, drawn without a clear picture of what went wrong, and then blended with feedback about something else entirely.
AgentGrad answers both with two components: sequential intervention, which fixes extraction, and semantic textual gradient abstraction, which fixes aggregation. The first decides where a correction belongs and what it should say. The second decides how corrections from many failures are combined into one prompt update.
For each failure, sequential intervention modifies the behaviour of one agent at a time, looking for the target agent whose modification resolves the failure. That settles the question earlier methods skipped, whether the chosen agent is really the one to blame. The modified output of that agent then becomes agent-level supervision, and a fine-grained gradient is extracted from it.
For aggregation, semantically similar gradients are clustered together so that unrelated failure modes are kept apart. Each cluster is then abstracted into one generalized gradient that captures the corrective pattern its members share.
AgentGrad achieves state-of-the-art performance across five MAS benchmarks. It also cuts wall-clock optimization time by two and a half times on average, compared with the next-fastest baseline. So the extra care in finding the right agent and grouping gradients by meaning does not come at a cost in time. On average it saves time.
Both fixes are about precision. The first makes sure feedback lands on the agent that caused the failure, backed by an example of what that agent should have produced. The second makes sure feedback from different kinds of failure is not blended into a prompt that serves none of them. What comes out is a prompt edited for the right reason and in terms general enough to carry beyond the cases that produced it.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A team of agents gets the answer wrong. Who do you fix, and how?
You run a team of language-model workers, each with its own role and its own instructions. When the team's final answer is wrong, you ask a model to write a critique and use it to revise an instruction.
The road ahead
- 1 · The problem ~7 min When a team of model workers gets a wrong answer, it is hard to tell whose instructions to revise, and pooled critiques muddle together.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Find the agent whose single fix repairs the failure, use its corrected output as a target, then group similar feedback and generalize it.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Per failure, hint agents one at a time to find the target, extract notes from corrected outputs, then cluster and abstract them.
- 6 · The experiment ~6 min You’ll guess: AgentGrad adds intervention runs and gradient clustering. How did its wall-clock optimization time compare with the next-fastest baseline?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.