Exploring Resource Allocation and Delivery Trade-offs in Deep Research

On November 11, 2026, a technical retrospective published on Juejin examined the trade-offs surrounding Deep Research, specifically in the areas of resource allocation, parallel evidence gathering, and delivery modes. Drawing on practices from OpenAI, Anthropic, Google, and the open-source community, the article analyzed how different deep research workloads demand fundamentally different approaches to compute budgets and final output formatting, using the ETF Agent engineering repository as a working example.

The piece opens by distinguishing between two broad categories of investment research tasks. Standard investment Q&A might require only verifying a single fund's fee rate. By contrast, another class of problems requires looking through portfolios, reading cross-period reports, handling revised versions, and coordinating different research branches before producing a coherent conclusion. The author points out that while both are called Research, they place entirely different demands on resource allocation and delivery pacing.

Rather than building a completely new agent for these heavy research tasks, the ETF Agent team took a different route. They allowed the same Native Research Loop to access higher, authorized resource tiers. Along with these elevated resources, the delivery guidance was adapted based on the active research mode: continuous progress feedback during the process, early delivery for sections with independent evidence, and more deliberate writing for key conclusions after a complete comparison of evidence was finished. The article notes that the standard Research loop already supported searching, reading, computing, calling Child agents, and delivering useful parts mid-task via send_answer_part. However, applying a uniform resource cap to all research meant complex tasks might start wrapping up before necessary evidence gathering was complete.

The author is explicit that the Deep implementation discussed is based on the ETF Agent engineering repository and does not claim these capabilities are already available in the current public Open ETF Agent. The article sets out to answer three main questions: why the original Research approach was insufficient, exactly what Deep changed at the code and runtime levels, and what results were delivered from actual Standard/Deep paired runs. To test these choices, the analysis leverages the ETF Agent implementation and six sets of paired experiments. The materials currently available do not include the specific results of these paired experiments.