Synthesis: What makes agent reason for longer
An agent can choose the wrong analysis, stop searching too soon, or retrieve the wrong records. We examine each failure mode through three 2026 studies, then look at related results from Foam’s coding and incident evaluations. These experiments cover different tasks and do not measure a common change in answer time.
Selecting an analysis
GeneBench-Pro evaluates 129 multistage computational biology problems in which an agent must inspect data, choose an analytical method, and revise its approach when diagnostics challenge its initial assumptions. In OpenAI’s evaluation, GPT-5.6 Sol passed 28.7% of the problems at its highest reasoning level.[1]
The benchmark scores the final scientific result. It does not measure the elapsed time required for ordinary agent questions. Its relevance here is the sequence of analytical decisions between receiving data and producing a supported answer.
Determining when search is complete
DeepSearchQA tests agents on 900 multistep, open-web research prompts across 17 fields, each requiring an exhaustive answer set. The authors identify two failure modes: stopping before all valid answers are found and adding low-confidence candidates that reduce precision.[2]
These errors have opposite effects on recall and precision. More searches alone cannot establish completeness; the agent must assess which parts of the answer set remain unverified.
Retrieving records under explicit constraints
VirBench evaluates 120 viral-sequence retrieval queries. Without a dedicated retrieval layer, the tested systems’ mean accuracy ranged from 16.9% to 91.3%. With gget virus, a deterministic interface for applying query constraints and retrieving complete result sets, every evaluated system reached at least 90% accuracy; runtime and tool calls generally fell.[3]
The intervention changed retrieval, not the underlying models. Its result is specific to a biological database, where structured filters and comprehensive pagination made the queries reproducible.
Evidence from coding and production incidents
Foam’s marginal tool utility study evaluated GPT-5.3-Codex on 25 public observability tasks. Removing Mattermost and Plane left one logs query tool; calls fell from 204 to 87 while task accuracy changed from 32% to 36%. The study measured calls and accuracy, not elapsed time or provider charges.[4]
In a separate set of 22 production incidents, Cursor identified the exact root cause in 9 cases with a Sentry tool and 14 with a Foam tool, using the same Cursor setup and model. This comparison measures diagnostic accuracy under two tool configurations. It does not establish the time or cost difference between them.[5]
Interpretation and limits
The external studies examine different agents and domains from Foam’s benchmarks. They expose different points of failure: choosing an analysis, knowing when a search is complete, and applying retrieval constraints. Foam’s results show fewer tool calls with a narrower tool set in one coding evaluation and higher root cause accuracy with the Foam tool in one incident evaluation. Neither result measures a general reduction in answer time.
A direct test would hold the agent and tasks constant, change one tool or data-access condition at a time, and report answer quality, tool calls, tokens, elapsed time, and provider charges for each condition.
References
- Jeremiah H. Li and Andrew J. Ho (OpenAI), GeneBench-Pro: Evaluating Multistage Statistical Reasoning in Genomics, Quantitative Biology, and Translational Biomedicine, 2026; see also OpenAI’s evaluation.
- Nikita Gupta et al. (Google DeepMind, Google Search, Kaggle, and Google Research), DeepSearchQA: Bridging the Comprehensiveness Gap for Deep Research Agents, 2026.
- Ferdous Nasri et al., Deterministic access to global viral sequence data enables robust agentic scientific discovery, 2026; discussed by Anthropic Research.
- Foam Research, Marginal tool utility in agents, 2026.
- Foam Research, Comparing agent harnesses and tools for incident investigation, with benchmark code and answer keys, 2026.