● Operational reliability for customer-support agents

Most agents already know the answer.
They fail on the process.

ResolveBench runs frontier AI agents through realistic, audited support workflows (authenticate, investigate, apply policy, execute the right tool, leave the database correct) and grades them on the actual end-state, not what they claim. Then it shows exactly why they fail. On the hardest cases, 57% of failures reached the correct, safe resolution and still failed, on a skipped check, a missing evidence ID, a prohibited action.

The core finding

Capability isn't the bottleneck. Discipline is

On the hardest cases, 57% of failing runs reached the correct, safe resolution, then failed the bar anyway: a skipped verification read, a missing evidence ID, a prohibited action. So we score the same 6,400 runs two ways. Outcome asks only "did the database end up correct?" Operational asks "correct and done by the book", the ResolveBench bar. The gap between them is reliability that single-run scores never see.

Outcome only: did the database end up correct? Operational: correct and by the book (pass⁸)
Claude Fable 5 high 81%52%
GPT-5.5 medium 73%39%
GPT-5.5 high 68%37%
Claude Opus 4.8 high 75%34%
GPT-5.6 Sol high 67%34%
Claude Opus 4.8 default 69%29%
Kimi K3 high 56%25%
Kimi K2.6 high 26%3%

Score on the final outcome alone and pass⁸ jumps by half to nearly double; every failure in that gap got the answer right but the process wrong. That gap, not the leaderboard, is what breaks agents in production.

The reliability gap

Every model looks better than it is

pass¹ asks "can it ever get this right?"; pass⁸ asks "does it get this right every time?" The drop between them is the reliability tax single-run benchmarks never charge.

Reliability decay: pass^k as k rises from 1 to 8. Hover any point for its value; click a model to toggle it.

pass¹: the chance a single run succeeds pass⁸: the chance all 8 runs succeed drop = reliability lost between one try and eight
Leaderboard

100 tasks · 8 trials each · pass^k reliability

Scored at a strict 90/100 composite bar and ranked by pass⁸: reliability, not average quality. Safety is a per-task hard-fail dimension. Reasoning effort is labeled per model: high runs think harder.

Score = mean composite quality 0–100 (a separate axis from reliability) · pass¹ = chance one run succeeds · pass⁸ = solves on all 8 runs · Safety = share of tasks with zero prohibited-tool use (n = tasks with a violation) · spark = pass¹→pass⁸ decay.

How agents fail

Four ways a capable agent still fails the case

Every run's full trajectory is public: golden plan beside the agent's actual tool calls. Across 6,400 runs the failures sort into four mechanisms, and almost none of them are "didn't know the answer."

01

Over-action under pressure

112 runs across 21 tasks fired a prohibited tool: escalating, refunding, or rebooking when the right move was simply to explain. Doing too much fails the case exactly like doing too little.

02

Right call, no evidence

The decision is correct but unsubstantiated. The typical miss cites just one of three required IDs; Evidence or tool-use is the weakest competence dimension for every model (best Evidence 4.17/5, Kimi K2.6 bottoms at 2.41).

03

Acting before looking

Skipped verification and authentication. 48.8% of runs miss a required tool call (vs 7.4% redundant), and a run that hits the auth gate without authenticating first fails 73% of the time.

04

Incomplete handoffs

The escalation decision is well-calibrated: 0 of 1,856 trials on escalation tasks substituted a forbidden action. What breaks is completeness: the handoff skips the verification reads and evidence a real one requires.

Failure decomposition

Each model fails its own way

Not all failures are equal. Here's what actually breaks each config's failing runs: the share caused by a wrong end-state, a missing evidence ID, skipped authentication, other tool-use errors, or an over-action. Almost everything is process, not the answer.

Why ResolveBench

Built for trust, not just a number

End-state
Scored on the database
Graded on the actual database state the agent leaves behind (what really changed), not a self-declared label.
Discipline
Operational compliance
Authenticate the caller, gather required evidence, take only allowed actions: the workflow is graded as part of the job, with prohibited tools a hard-fail traced to the exact step.
Transparency
Every run is public
Golden plan beside the agent's actual tool calls, for every run, with the exact failure reason. The whole benchmark is auditable: all 6,400 runs are on Hugging Face ↗.
pass^k
Reliability, not luck
Every task runs 8× and we report the full decay curve, so a flashy one-shot score can't hide run-to-run variance.
Coverage

Three domains today. Many more on the way.

The published benchmark covers airline, hotel, and utility support. The methodology is domain-agnostic: the same end-state scoring, operational checks, and audited goldens apply anywhere, so we're expanding into the workflows enterprises ask about most.

Live now
AirlinesHotelsUtilities
Expanding soon
BankingInsuranceHealthcare admin TelecomE-commerceSaaS support HR operationsTravel booking
FAQ

Common questions

What is ResolveBench?
ResolveBench is an operational-reliability benchmark for customer-support AI agents. It runs frontier agents through realistic, audited support workflows and grades them on the actual database end-state they leave behind (plus whether they authenticated the caller, gathered the required evidence, and avoided prohibited actions) across 8 independent runs per task. Every run's golden-vs-agent trajectory is public, so you can see exactly why an agent failed.
What's the single most important finding?
Capability isn't the bottleneck; operational discipline is. On the hardest cases, 57% of failing runs reached the correct, safe resolution and still failed, on a skipped verification read, a missing evidence ID, or a prohibited action. Re-scoring on the final outcome alone raises pass⁸ by half to nearly double: most failures got the answer right but the process wrong.
What is pass^k (and pass⁸)?
pass^k is the probability that an agent succeeds on all k independent attempts at a task. pass^1 asks 'how often does a single run succeed?'; pass^8 asks 'does it solve it every single time?' The gap between them is the reliability that single-run benchmarks miss.
Which AI model is most reliable on ResolveBench?
Claude Fable 5 (high effort) leads outright with pass^8 of 52%, a statistically significant lead (paired bootstrap +13 points over the runner-up, 95% CI [+2, +25]). Behind it, GPT-5.5 medium (39%), GPT-5.5 high (37%), Claude Opus 4.8 high (34%), and GPT-5.6 Sol (34%) form a statistical tie; Claude Opus 4.8 default posts 29%, Kimi K3 (25%) lands just below it with overlapping confidence intervals, and only Kimi K2.6 (3%) separates at the bottom. Within the GPT-5.5 family, more reasoning effort still did not produce a measurable reliability gain.
How is ResolveBench different from τ-bench?
τ-bench largely measures whether an agent can complete a task. ResolveBench measures whether it completes the task the way a compliant support employee must: graded on the actual database end-state (not a self-declared label), with authentication, evidence, and allowed-action checks folded into the score, a per-task safety hard-fail, and every run's full trajectory published. Its focus is failure decomposition: was the failure procedural or reasoning-related? Both benchmarks run tasks multiple times; the difference is operational reliability and transparency, not the number of trials.
How many tasks, models, and domains does it cover?
100 audited cases across three live domains (airline, hotel, and utility support) run 8× each across 8 frontier model configurations (6,400 scored trajectories), all open for inspection. More domains (banking, insurance, healthcare administration, telecom, e-commerce, and more) are on the way.
Can I evaluate my own model or domain?
Yes. ResolveBench builds audited tasks for your support workflows and evaluates any model, including fine-tunes, under the same methodology. Request custom tasks via the contact page.
Who built ResolveBench?
ResolveBench is an open benchmark released by Hubble42 Inc.

Want your model, or your domain, on this board?

We build audited tasks for your support workflows, evaluate any model on them, and help you close the reliability gap.

Request custom tasks