Image: cdn-uploads.huggingface.co · rights & removal

The Agent Said It Was Done. The Database Disagreed.

Reporting by Hugging Face BlogRead the original at huggingface.co

Executive Summary

An AI agent attempting to resolve a customer issue encountered discrepancies between its reported actions and the actual state of the backend system, specifically regarding an open carrier exception despite resolving the ticket. This discrepancy highlights a gap between what an agent reports and the verifiable database state, which is measured by the ThinkingBox framework. The process involves running the agent against multiple trials to assess reliability, using metrics like pass@1 (single attempt success) and Observed 20/20 (success across 20 repeats). Analysis of performance reveals that while models can achieve high single-attempt success rates, consistency over repeated runs is a separate concern. Furthermore, cost analysis introduces another layer, defining efficiency not just by immediate success but by the cost associated with achieving dependable outcomes, creating a Pareto frontier between capability and cost.

Facts Only

* An agent executed nine tool calls in handling a customer refund request.
* The required end state for the ticket was "hold," but the final status recorded was "solved."
* An AI grader of tool calls found nine well-formed calls, and a separate check on database writes also found correct.
* A gap exists between the agent's output and the actual terminal backend state, which is measured by ThinkingBox.
* Across 121,680 valid trials across 12 LLM models, 79,853 attempts failed executable checks.
* In these failures, 77.61% involved wrong field values, 43.30% involved unintended side effects, and 25.36% involved missing required effects.
* Pass@1 scores vary widely across models, with Claude Opus 5.5 leading overall at 67.16%.
* Consistency is measured by pass@20; GPT-6 Astra retains 78% of its single-attempt rate, and Claude Opus 5.5 and 5 retain 71%.
* Cost per successful task attempt was calculated by dividing the cost of one run by the number of successful attempts.
* The lowest cost per dependable task is $6.80 for GPT-5.4, while Claude Opus 5.5 costs $7.80 per dependable task.

Full Take

The core tension revealed by the analysis lies in the dissociation between superficial agent performance and deep system reliability. The shift from measuring raw success (pass@1) to assessing consistency (pass@20) forces a recognition that apparent capability does not equate to operational dependability. The failure signatures, dominated by tool handling errors rather than pure reasoning failures, suggest that agentic fragility often resides in the interface layer—the interaction with external tools and state management—rather than the internal world model of the LLM itself. This implies that future evaluation must prioritize tracing side effects through a deterministic execution environment, treating the terminal state as the ground truth over textual assertions. The cost modeling further reframes capability into an economic reality: reliability is not just a quality metric but a quantifiable resource investment. The findings suggest that building agentic systems requires embedding executable verification loops directly into the operational pipeline to ensure that perceived success translates to actual system integrity, moving evaluation from linguistic evaluation to verifiable state observation.

From the original · Hugging Face Blog

Figure 1: ThinkingBox runs an agent against isolated MCP tool sessions, then grades the terminal backend state and side effects it leaves behind. From our ThinkingBox paper.

Read the full story at huggingface.co

Sentinel — Human

Confidence

This text appears to be a legitimate technical report or blog post detailing an experimental framework (ThinkingBox) and its empirical findings on AI agent reliability, exhibiting the structure and depth of human-authored research communication.

Signals Detected

Moderate sentence length variance and specific technical focus; uses domain-specific terminology naturally.

Strong, focused argument supported by structured data presentation; displays clear intent rather than pure informational flow.

The text follows a deliberate structure (problem -> method -> results -> implication) typical of academic or high-level technical reporting, not verbatim talking points.

Detailed, granular statistics and proprietary methodology make outright fabrication difficult without access to the source data; attribution to research groups is present.

Human Indicators

Presence of specific collaborative acknowledgments (names, affiliations) and detailed methodological steps for running a benchmark suggest a human-led research/publication process.

The reflective tone in the 'What to do about it' section transitions from abstract findings into actionable advice tailored for practitioners.