Agentic AI
New Framework Helps AI Agents Search Long, Complex Tasks Without Running Out of Context
Researchers introduce SLIM, a lightweight system that lets agentic search agents handle longer research chains by managing memory and tool use more efficiently.
Agentic search systems—AI agents that explore the web iteratively to answer complex research questions—hit a fundamental wall: they quickly consume their available context window with accumulated clutter and exhaust their tool budgets before finishing long tasks. A new research paper presents a solution designed around simplicity and focus.
The Core Problem#
As AI agents conduct multi-step research across numerous sources, they accumulate noisy browsing history, hit context limits, and run out of allowed tool calls. This forces them to cut searches short or miss key information. The friction isn't from the underlying models—it's from how search and information retrieval are organized.
SLIM: A Lightweight Approach#
Researchers introduce SLIM (Simple Lightweight Information Management), which reimagines the agent's workflow through two key changes:
- Separated search and browse: Instead of mixing exploration with content extraction, SLIM cleanly divides these into distinct tools
- Periodic summarization: The system compresses trajectory history at intervals, keeping context tidy while preserving what matters
The framework is deliberately minimal—no complex memory systems or elaborate state management.
Results Across Benchmarks#
Testing on long-horizon research tasks shows substantial gains:
- With OpenAI's o3 model, SLIM reaches 56% accuracy on BrowseComp and 33% on HLE, beating competing open-source frameworks by 8 and 6 points
- It uses 4–6 times fewer tool calls than other systems, reducing computational cost significantly
- On GLM-4.7 Flash, SLIM outperforms the next-best framework (Search-o1) by 10 points on BrowseComp while consuming only one-third the resources
Understanding Where Systems Fail#
The researchers developed an automated error-analysis pipeline and taxonomy to diagnose failure modes. They found that SLIM exhibits notably fewer hallucinations—a critical problem in agentic search where false claims can derail entire chains of reasoning. This systematic breakdown of errors offers insights into how long-horizon agents can be made more reliable.
