SAgEScience of Agent Evaluation Emerging Trends in AI Agent Evaluation
SAgEScience of Agent Evaluation

Emerging Trends in AI Agent Evaluation

Felix Chen1,* Stephan Rabanser1,* Peter Kirgis1 Nitya Nadgir1 Rishi Bommasani2 Magda Dubois3 Elham Tabassi4 Cozmin Ududec3 Anka Reuel2 Sanmi Koyejo2 Sayash Kapoor1,5,* Arvind Narayanan1,6,*

1Princeton University 2Stanford University 3UK AI Security Institute 4Brookings Institution 5UC Berkeley 6Institute for Advanced Study

*Corresponding authors: {felix_chen, rabanser, arvindn}@princeton.edu, sayashk@berkeley.edu

Read the paper Explore the 33 trends Cite

Overview

The how, what, and who of agent evaluation

Questions of interest for evaluation of AI agents today no longer implicate models alone, but rather expand to the larger systems around them, consisting of scaffolds (also referred to as “agent harnesses”), tools, testing environments, and user interfaces. As a result, long-standing concerns may manifest differently for agents than they did for models, and previous assumptions may not hold.

How? Rethink evaluation methodology An agent’s score depends on its scaffold and environment as well as its model. Long, stochastic trajectories need better statistics, tools open new routes to contamination, and agents can exploit bugs in tasks and graders. Part 1 · 13 trends · best practices What? Broaden the scope of evaluation Evaluation is moving from the lab to the field, where agents are now deployed. Beyond what an agent can do alone, evaluations must measure how it works with people and how it changes the organizations that use it. Part 2 · 11 trends · research directions Who? Recognize the role of evaluators Who runs an evaluation shapes its result. The same model and benchmark can score differently depending on who chooses the scaffold, prompts, tools, environment, compute budget, and statistical analysis. Part 3 · 9 trends · practical considerations

Model evaluation and agent evaluation

Model evaluation grades a single response against a reference answer. Agent evaluation covers a system of model, scaffold, and tools that acts in an environment over many steps, sometimes with a human stepping in. Instead of a reference answer, the grader uses tests, a rubric, or an LLM judge to check the trajectory (the process), the final state (the result), or both.

Swipe sideways to see the whole diagram.

Conclusion

There is a clear window now to bring in new participants, methods, and ideas for agent evaluations, and to meet the pressing need for rigorous evidence in a fast moving space.

Takeaways

All takeaways

The best practices, research directions, and practical considerations from every trend.

Terms

Definitions

References

All references

About

SAgE Science of Agent Evaluation

The SAgE group at Princeton University works on advancing the systematic study and evaluation of agents.

Corresponding authors

    Visit the SAgE website

    Citation

    @misc{chen2026emerging,
  title  = {Emerging Trends in AI Agent Evaluation},
  author = {Chen, Felix and Rabanser, Stephan and Kirgis, Peter and Nadgir, Nitya and Bommasani, Rishi and Dubois, Magda and Tabassi, Elham and Ududec, Cozmin and Reuel, Anka and Koyejo, Sanmi and Kapoor, Sayash and Narayanan, Arvind},
  year   = {2026}
}

    Acknowledgements