Evaluation of Agentic AI Systems Under Realistic Constraints

Establish reproducible and realistic evaluation frameworks and benchmarks for agentic artificial intelligence systems used across scientific workflows that accurately measure performance under real-world constraints, including adherence to provenance, claim-level citation correctness, tool-call success rates, inter-agent handoff quality, confidence calibration, and time or cost to solution when facing API failures, shifting corpora, and strict latency or compute budgets.

Background

The paper argues that agentic AI systems—those that plan, act, and reflect across stages of scientific work—are increasingly used from literature retrieval to experiment design and execution. Despite this growing role, the authors note that assessing such systems in conditions that mirror real scientific practice is unresolved.

They highlight that current testbeds reveal issues such as data contamination, brittle scaling with compute, and poor generalization beyond memorization. The authors suggest that evaluation must go beyond task accuracy to include process-oriented metrics and robustness to operational constraints commonly encountered in research environments.

References

Evaluating agentic systems under the constraints they face in practice remains an open problem.

— Rethinking Science in the Age of Artificial Intelligence  (2511.10524 - Eren et al., 13 Nov 2025) in Subsection "Realistic Benchmarks and Evaluation" under Section "AI and Science"

Second, the current evaluation considers two canonical scenarios; the paradigm's performance under more heterogeneous deployments involving satellite-terrestrial integration, UAV-assisted networks, and industrial IoT environments remains an open question.

— ASTRA: Toward Agentic AI for Intelligent Device-Network-Cloud Synergy in Next-Generation Mobile Communication  (2609.21298 - Guo et al., 18 Sep 2026) in Section Conclusion

What the fix does not bound is the grading surface itself, since the confidence floor of 0.6 on three ACS70331 claims and the required-substring needles were fixed inside the leak window and never revisited. We cannot rule out that leaked runs shaped them, which is why Appendix~\ref{sec:appendix-rederivation} re-derives the needles blind (its floor arm was lost to ambiguous instructions).

— Fidelity Is Not Enough: Dispatch-Level Instrumentation for Agentic Datasheet Extraction  (2608.28439 - Ye et al., 28 Aug 2026) in Section 6, paragraph “What the fix did not touch” (Section~\ref{sec:appendix-audit})

While such long-horizon evaluation is critical for sustained capability growth, current studies primarily target software or web-based tasks, leaving open the question of how similar principles can be applied to complex EO workflows with evolving environmental conditions and multi-sensor inputs.

— RingMoClaw: An Experience-Inspired Multi-Agent Framework for Self-Evolving Research in Remote Sensing  (2609.00814 - Kang et al., 1 Sep 2026) in Section 2.3, OpenClaw-Based General Agent Frameworks

Evaluating changing membership and cross-domain trust remains an open benchmark problem.

— SoK: When Safe Agents Fail Together: The Security of Multi Agent LLM Systems  (2609.00595 - Yang et al., 1 Sep 2026) in Section 6, subsection 6.5, “Open-System Evaluation”

Finally, our evidence comes from A/B experiments that measure real effects but are costly to run and specific to their setting; a standardized way to evaluate agent-driven system optimization before deployment remains an open problem, and one we hope this work helps motivate.

— CORAL: An LLM-Native Harness for Production Recommender Systems  (2609.02730 - Azhar et al., 2 Sep 2026) in Section Conclusion

The agentic toolqa family directly addresses the not-really-agentic objection this design choice invites, but broader ecological validity against fully open-ended real-world agentic workflows remains an open question beyond what this study can settle.

— How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making  (2609.01660 - Mittal, 31 Aug 2026) in Section 5, Limitations and Threats to Validity

Template test cases with runtime-resolved placeholders provide reusable regression coverage across the evaluated specifications, models, and harnesses; longitudinal validation under actual API evolution remains future work.

— Continuous Process-Level Evaluation for Evolving Enterprise AI Agent Skills  (2610.01833 - Vo et al., 1 Oct 2026) in Abstract; Section 3, paragraph on CI/CD-style regression evaluation

Productivity extends Revenue with external cost sources, cost allocation, NULL handling, and a different persistence path; generalization to other skill types, tool suites, or non-quantitative outputs remains open.

— Continuous Process-Level Evaluation for Evolving Enterprise AI Agent Skills  (2610.01833 - Vo et al., 1 Oct 2026) in Appendix, Section "Limitations"

It is also unclear what successful behaviour looks like, since ``doing good science'' is a fuzzy construct , and the final artefact is a written paper or report, whose quality may conceal many issues during its construction, including data fabrication, citation hallucination, and deep conceptual misunderstandings .

— Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations  (2610.08364 - Pilditch et al., 6 Oct 2026) in Section 3, “Case Study: Reading an AI R&D Evaluation with Transect,” opening paragraph