Agent Evaluation Benchmarks

The evaluation landscape for LLM agents has matured rapidly, growing from isolated function-call correctness tests to comprehensive interactive benchmarks. Ferrag et al. consolidate approximately 60 benchmarks developed between 2019 and 2025 into a unified taxonomy spanning 8 domains: general knowledge reasoning, mathematical problem-solving, code generation, factual grounding, multimodal tasks, and interactive assessments.

Hu et al. propose a complementary three-tier evaluation framework specific to tool use: tool-call validity (does the model produce syntactically correct tool invocations?), task completion (does the tool chain achieve the goal?), and interactive performance (how well does the agent perform in realistic, multi-turn environments like WebArena and OSWorld?).

A critical finding across both surveys: benchmark performance does not reliably predict real-world deployment robustness. This gap represents one of the field's most pressing challenges.

Longitudinal Capability Trends (AISI, Dec 2025)

Beyond static benchmark snapshots, the UK AI Security Institute's first Frontier AI Trends Report (Dec 2025) supplies the field's most rigorous time-series of frontier capability, drawn from controlled lab evaluations of 30+ models across 2022–2025. The headline pattern is a measurable, accelerating doubling rate in security-relevant task horizons:

• Cyber (Cybench-style): apprentice-level task success rose from ~9–10% (late 2023) to ~50% (current); the length of cyber tasks models can complete unassisted is doubling roughly every 8 months — and recent (Feb 2026) internal data put that at ~4.7 months since late 2024.
• Software engineering: ~one-hour SWE tasks succeeded >40% of the time (summer 2025).
• Autonomy (RepliBench): self-replication task success rose from <5% (early 2023) to >60% (mid-2025) across five frontier models — though confined to controlled, simplified environments.

These trends matter for evaluation because they show capability progress is best read as a rate (doubling time), not a fixed score — and that the rate is itself accelerating.

The Matched-Budget Correction (Wang et al., Jul 2026)

The sharpest 2026 contribution to evaluation methodology is not a new benchmark but a control. Wang et al. (arXiv:2607.12227) show that a large class of reported agentic gains — those from automatic harness evolution, i.e. searching over scaffold/tool configurations — are confounded in two ways that generalize well beyond harness work:

1. Budget confound. Any method that iteratively evaluates and revises candidates using task feedback is spending inference compute. Unless it is compared against a simple task-level search baseline under matched feedback and inference budget, an improvement cannot be attributed to the method rather than to the extra compute it consumed.
2. Search/eval benchmark sharing. When the search loop and the reported evaluation use the same public benchmark, the headline number is partly a measure of overfitting to that task set. The corrective is a held-out task split.

Applied to Terminal-Bench 2.1 with GPT-5.4 and Claude Opus 4.6, harness evolution does not consistently beat simple test-time scaling and generalizes poorly to held-out tasks. Full treatment at Agent Harness Evolution.

The general lesson for this page: matched budget and held-out evaluation should now be the default questions asked of any agentic result, including vendor-published ones — which, as catalogued in Inference & Token Economics, essentially never disclose search or inference budget at all.

Two more evaluation confounds the July pass measured: the harness, and the budget

The matched-budget correction has company. Three harness papers and one inference-compute paper (all ingested 2026-07-23) each name a different uncontrolled variable that current agent leaderboards hold implicit — and each shows it moving the number.

The harness is an undisclosed variable that can dominate the model. Stop Comparing LLM Agents Without Disclosing the Harness measures harness-induced variance at 7.80× model-induced variance (18.48 vs 2.37 pp², 6/9 ranking reversals) in a controlled 3×3 SWE-bench factorial, and argues that long-horizon agent leaderboards are "incomplete and potentially misleading" until the harness is disclosed — proposing an ETCSOVG Harness Card and a variance-decomposition reporting protocol. Harness-Bench supplies the diagnostic instrument (106 tasks, 5,194 trajectories) and shows performance must be attributed to model-harness configurations, not models. Belief-divergence shows that even outcome-matched harnesses drive different mid-trajectory beliefs — so pass/fail parity is not evidence of harness equivalence. Full treatment on Agent Harness Evolution; the evaluation lesson here is that "which model scored X" is unanswerable without also stating the harness, and no major leaderboard yet requires it.

Inference budget is an evaluation axis, not a hidden constant. How Inference Compute Shapes Frontier LLM Evaluation (up to 12 frontier models × 7 benchmarks, controlled inference-scaling interventions: token-budget expansion, context compaction, repeated attempts) shows that fixed single-budget scores increasingly understate models — and the underestimation grows for newer models, which have a higher ceiling that only opens at generous budgets. Its recommendation is to report capability as a function of inference-time compute rather than as a point estimate. This is the other face of the matched-budget correction: Wang et al. warn that unequal budgets inflate a method's apparent gain; McFadyen et al. warn that a fixed low budget deflates a model's apparent ceiling. Together they make inference budget a first-class evaluation parameter — one that has to be swept and reported, not fixed and hidden.

The combined standard this page now carries: an agentic result is only interpretable if it discloses (1) the harness, (2) the search/inference budget, and (3) held-out generalization — and reports capability across a budget range, not at one point.

Key Claims

• ~60 benchmarks exist across 8 evaluation domains — Developed 2019-2025, covering reasoning, math, code, factual, multimodal, interactive. Evidence: strong (From LLM Reasoning to Autonomous Agents)
• Three-tier evaluation framework for tool use — Validity → completion → interactive performance. Evidence: strong (Agentic Tool Use in LLMs)
• Benchmark-to-deployment gap persists — Performance on benchmarks doesn't fully transfer to real-world scenarios. Evidence: moderate (From LLM Reasoning to Autonomous Agents)
• Cyber-task time horizon is doubling every ~8 months, accelerating to ~4.7 months — AISI's longitudinal evaluation of 30+ models (2022–2025); apprentice-level success 9%→50%. Evidence: strong (AISI Frontier AI Trends Report)
• Self-replication (RepliBench) success rose <5%→>60% (2023→mid-2025) — but only in controlled, simplified settings; models are weakest at later stages (persistent access). Evidence: strong (AISI Frontier AI Trends Report)
• Gains from iterative search methods are unattributable without matched-budget baselines — Harness evolution is structurally a test-time-scaling method; unless feedback and inference budget are equalized against a simple search baseline, "better method" and "more compute" cannot be separated. The argument is analytic and applies to any feedback-driven agentic method, not only harness search. Evidence: strong (Rethinking the Evaluation of Harness Evolution)
• Searching and reporting on the same benchmark inflates results; held-out splits are required — On Terminal-Bench 2.1 (GPT-5.4, Claude Opus 4.6), evolved harnesses showed limited generalization to held-out tasks, indicating the search loop overfits the task set it optimizes against. Evidence: strong (Rethinking the Evaluation of Harness Evolution)
• The harness is an undisclosed evaluation variable that can dominate the model — Harness-induced variance measured at 7.80× model-induced variance (18.48 vs 2.37 pp², 6/9 ranking reversals) on a controlled SWE-bench factorial; leaderboards without harness disclosure are "incomplete and potentially misleading." Evidence: moderate (controlled factorial, small n, self-measured; position paper) (Stop Comparing, Harness-Bench)
• Fixed single-budget scoring understates newer models; capability should be reported across an inference-budget range — Up to 12 frontier models × 7 benchmarks under controlled inference-scaling interventions; underestimation grows for newer models with higher ceilings. Inference budget is an evaluation axis, not a hidden constant. Evidence: moderate-to-strong (preprint, third-party, methodologically careful; numeric grid not captured) (How Inference Compute Shapes Frontier LLM Evaluation)

Benchmarks & Data

• ~60 benchmarks taxonomized (Ferrag et al.)
• Key interactive benchmarks: WebArena, OSWorld (Hu et al.)
• 11 real-world application sectors documented (Ferrag et al.)
• Terminal-Bench 2.1 used as the agentic testbed for the matched-budget correction (Wang et al.)
• Harness variance : model variance = 7.80× on a 3×3 SWE-bench Verified factorial (Stop-Comparing); 106 tasks / 5,194 trajectories isolating harness effects (Harness-Bench)
• Inference-budget sweep across up to 12 frontier models × 7 benchmarks (SWE, math, medicine, cyber) shows fixed-budget scores understate newer models (McFadyen et al.)

Open Questions

• How to build benchmarks that reliably predict real-world agent performance?
• Can evaluation keep pace with rapidly expanding agent capabilities?
• How to benchmark multi-agent collaborative scenarios fairly?
• What metrics capture safety failures, not just capability?
• How many previously published agentic gains survive a matched-budget, held-out re-evaluation? Nobody has swept the literature with this control yet.
• How should the matched-budget standard be enforced against vendor claims, where search and inference budgets are never disclosed?

Related Concepts

• Agentic Reasoning — Benchmarks measure agentic capabilities
• LLM Tool Use — Three-tier evaluation framework specific to tool use
• Multi-Agent Systems — Interactive benchmarks increasingly test collaboration
• Agent Safety & Alignment — The AISI report's biorisk/jailbreak findings sit on the safety side
• Agent Harness Evolution — the specific literature the matched-budget correction was aimed at; full home of the 7.80× harness-variance finding and the harness-disclosure cluster
• Inference & Token Economics — where "capability is a function of inference budget" becomes a per-task pricing argument

Backlinks

Pages that reference this concept:

• OpenAI
• DeepSeek
• Agent Harness Evolution

Changelog

• 2026-06-24 — Compiled new source (aisi-frontier-ai-trends-report-2025): added longitudinal capability-trend section (Cybench doubling ~8mo→~4.7mo, RepliBench <5%→>60%, 1hr SWE >40%)
• 2026-07-22 — Compiled new source (rethinking-harness-evolution-evaluation): added the matched-budget / held-out-split correction as a general evaluation standard, two claims, and open questions on re-evaluating prior results and vendor claims; frontier_status steady → active
• 2026-07-23 — Compiled four evaluation-methodology sources. Added a "Two more evaluation confounds" section: the harness as an undisclosed variable (Stop-Comparing 7.80×, Harness-Bench 5,194 trajectories, belief-divergence) and inference budget as a first-class evaluation axis (How Inference Compute Shapes Frontier LLM Evaluation — fixed-budget scoring understates newer models). Consolidated the emerging standard: disclose harness + budget + held-out, report capability across a budget range. Two new key claims and two benchmark-data rows. Deep treatments stay on Agent Harness Evolution and Inference & Token Economics.