TLDR: AgentFold is a new AI agent paradigm that improves performance on long and complex web tasks by intelligently managing its working memory. Inspired by human cognition, it uses ‘folding’ operations—either condensing recent steps or consolidating entire sub-tasks—to keep its context concise and focused, preventing information overload or loss. This allows it to outperform larger models and proprietary agents on benchmarks, demonstrating exceptional efficiency and scalability for extended web interactions.
In the rapidly evolving world of artificial intelligence, large language model (LLM)-based web agents hold immense potential for tasks like information gathering and synthesis. However, these agents often hit a wall when faced with long and complex tasks. The core issue lies in how they manage information, or ‘context’. Traditional methods either accumulate too much raw, noisy data, leading to confusion, or they summarize everything too aggressively, risking the loss of crucial details.
A new research paper introduces an innovative solution called AgentFold, a novel agent paradigm that tackles this problem head-on. Inspired by how humans consolidate memories and manage their thoughts, AgentFold treats its working memory not as a passive log, but as a dynamic workspace that it actively sculpts. This proactive approach allows it to maintain a focused and efficient context, even during extended interactions.
At the heart of AgentFold is its unique ‘folding’ operation, which manages historical information at multiple levels. Imagine an agent working through a task; instead of just adding every new piece of information to a growing list, AgentFold intelligently decides what to keep and how to store it. It can perform ‘granular condensations’ to preserve important, fine-grained details from recent steps, ensuring nothing vital is overlooked. Alternatively, it can execute ‘deep consolidations’ to abstract away entire multi-step sub-tasks once they are completed, effectively packaging a long sequence of actions into a single, concise conclusion. This prevents the agent’s memory from becoming cluttered with irrelevant intermediate steps.
AgentFold’s context is structured into four key components: the original user question, a list of available tools, ‘Multi-Scale State Summaries’ (acting as its long-term memory of distilled past events), and the ‘Latest Interaction’ (its immediate working memory of the most recent action and observation). This design ensures the agent always has access to both a high-level overview and the most current details.
The agent’s operational cycle involves a continuous loop of perceiving, reasoning, folding, and acting. After each action, AgentFold generates a response that includes a folding directive, an explanation of its thought process, and the next action. This integration of context management as a core, learned step is what sets AgentFold apart, allowing it to dynamically balance the need for detailed information with the necessity of a coherent long-term plan.
To train this intelligent agent, the researchers developed ‘Fold-Generator’, a specialized data collection pipeline. This pipeline automatically creates detailed interaction trajectories that demonstrate sophisticated context curation, which are then used to fine-tune open-source LLMs. This supervised fine-tuning process teaches the AgentFold model to intrinsically produce its structured outputs, making it robust and efficient.
The results of AgentFold are impressive. The AgentFold-30B-A3B model achieved state-of-the-art performance on prominent information-seeking benchmarks like BrowseComp and BrowseComp-ZH, and general benchmarks like WideSearch and GAIA. Notably, it outperformed open-source models significantly larger in size (up to 20 times larger) and even matched or surpassed leading proprietary agents like OpenAI’s o4-mini. This demonstrates that effective context management can bridge the performance gap with much larger models.
Further analysis revealed AgentFold’s exceptional context efficiency. Its average token count grew at a remarkably slow, sub-linear rate, reaching only about 7,000 tokens after 100 turns, a small fraction of its underlying model’s capacity. This efficiency allows AgentFold to scale to hundreds of interaction steps, a critical capability for truly long-horizon tasks that often overwhelm conventional agents. The ability to perform deep consolidation also allows AgentFold to recover from dead ends by abstracting away unsuccessful lines of inquiry, resetting its context to a more compact state and pivoting to new strategies.
Also Read:
- Boosting Web Agent Reliability with Confidence-Guided Retries
- ATLAS: A New Web Agent That Learns and Plans Through Simulated Action
In conclusion, AgentFold represents a significant leap forward in AI agent design. By empowering agents to act as self-aware knowledge managers, it overcomes fundamental limitations in context handling, paving the way for more robust, efficient, and scalable AI systems capable of tackling complex, long-duration web tasks. For more details, you can read the full research paper here.
