Tech Daily Report: AI Agent Ecosystem & Research Knowledge Systems (2026-09-12)

1. Multi-Agent & Deep Research Papers (ArXiv)

Mr.LHDR: A Benchmark for Multimodal Real-World Long-Horizon Deep Research Agents

• URL: https://arxiv.org/abs/2609.11318
• Published: 2026-09-10
• Summary: Introduces a benchmark focusing on "long-horizon" deep research over irreducible chains of interdependent evidence. It evaluates whether agents can sustain long reasoning processes. Findings show even the strongest systems achieve only 43.1% Overall Accuracy (OA) and 34.3% Strict Accuracy (SA), showing that sustaining dependency-consistent evidence integration over time is a key bottleneck.

Benchmarking Hybrid Deep Research Across Database Querying and Web Search

• URL: https://arxiv.org/abs/2609.09410
• Published: 2026-09-08
• Summary: Introduces HybridDeepResearch, the first benchmark requiring agents to weave evidence from both unstructured text (open web) and structured data (relational DBs/SQL) to form complete answers. Models like Claude-Sonnet-4.6 and GPT-5 achieve ~50-54% Pass@8 on hard tasks, showing that bridging unstructured and structured information without losing constraints is difficult.
• Repository: https://github.com/Snowflake-AI-Research/HybridDeepResearch

When Agents Disagree: Bayesian Backward Reasoning as a Label-Free Anchor for Multi-Agent Collective Decision-Making

• URL: https://arxiv.org/abs/2609.11709
• Published: 2026-09-10
• Summary: A novel method for multi-agent collective decision making using Bayesian backward reasoning when multiple agents yield conflicting answers. Offers better cross-path consistency to aggregate agents' outputs over standard voting methods.

ORCH: Organizational Principles Enable Collective Intelligence in Embodied AI

• URL: https://arxiv.org/abs/2609.11737
• Published: 2026-09-10
• Summary: Constructs task-specific hierarchical organizations (ORCH) combining pooled interdependence for concurrent work with sequential interdependence for work with prerequisite relationships. Showcases a 63.97% improvement in final score and 74.29% execution efficiency across missions compared to prior multi-agent frameworks.

FrankenReport: Early Exiting in Long-Form Generation Using Expected Value of Computation

• URL: https://arxiv.org/abs/2609.06320
• Published: 2026-09-06
• Summary: Presents an adaptive early exiting method for long-form knowledge-seeking report generation, assessing intermediate outputs to predict if further computation yields quality gains, improving resource consumption.

2. Open-Source Repositories & Agent Frameworks (GitHub)

letta-ai/letta (f.k.a. MemGPT)

• URL: https://github.com/letta-ai/letta
• Description: Platform for stateful agents with advanced, long-term memory that can learn and self-improve over time.
• Role: Addresses the long-term memory challenge in multi-agent orchestration for deep research and information retrieval.

Snowflake-AI-Research/HybridDeepResearch

• URL: https://github.com/Snowflake-AI-Research/HybridDeepResearch
• Description: Code and datasets for Hybrid Deep Research over Web and SQL data. Contains implementations of Web Search, SQL execution, and knowledge tools.
• Analysis: Running a graft build on this repository reveals 14 core files and 84 functions. Core capabilities involve evaluation/execution_match.py (assessing structural and dataframe similarities) and tools/ providing search/SQL integrations (tools/run_sql.py, tools/web_fetch.py, tools/knowledge.py).

assafelovic/gpt-researcher

• URL: https://github.com/assafelovic/gpt-researcher
• Description: An autonomous agent that conducts deep research on any data using any LLM provider, providing a functional, production-ready implementation of automated literature review and synthesis.

3. Graft Codebase Analysis (HybridDeepResearch)

A graft map was generated for the Snowflake HybridDeepResearch repo: - Scope: 14 files, 84 symbols, 235 edges. - Hubs: - evaluation/execution_match.py (_columns_equal, _dataframe_similarity, _rows_for_dataframe) - heavily focused on robust comparisons. - tools/knowledge.py (_kb_path, _load_entries) - handles knowledge-base path mapping and resolution. - preprocess/build_dataset_releases.py (with_task_type, build_releases) - structured formatting. - Conclusion: The codebase is lean and focuses strictly on accurate tooling execution and complex evaluation over structured and unstructured contexts, reflecting the need for tight logic integration.