跪拜 Guibai
← All articles
Artificial Intelligence · Programmer

Agent Demos Are One Lucky Sample; Production Runs the Whole Distribution

By 深蓝AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Anyone picking an agent by leaderboard score is betting on a cherry-picked sample. Evaluation infrastructure now decides agent reliability costs more than model choice does, and teams that skip it will discover the 1% tail only after it has already been billing them.

Summary

Public leaderboards measure an agent's best-case ceiling: Claude Fable 5.2 completes 82.9 LiveBench 2.0 tasks, yet the best model in EON's simulated enterprise environment, Claude Opus 4.8, only clears 76.8% of tasks. The difference is a structural feature of how agents are evaluated, not a defect of any single model. A demo records the one perfect pass out of ten runs and cuts the failures, but production customers experience every run, so a 1% aberration becomes ten thousand daily incidents. Nobody measures the floor.

TechCrunch's November 11 report catalogs the repeated failure modes: strategy drift, where an agent hits the task but ignores a budget policy; multi-source entanglement, where combining three datasets stalls a flow; and long conversations that crush the context window. These compound multiplicatively when combined.

Plurai, with a $13M seed led by Team8, replicates your company's systems as a digital twin, maps every agent action path into a graph, and systematically walks the edge cases, then post-trains a custom small open-source model for each client from the simulation data. The practical floor for any team: twenty real task flows, three injected edge types, and a Wilson confidence interval on every reported success rate.

Takeaways
— LiveBench 2.0 results: Claude Fable 5.2 completes 82.9 general tasks and ChatGPT-6 Astra completes 75.6; in EON's simulated enterprise environment the best model, Claude Opus 4.8, tops out at 76.8% task success.
— A demo is a selected single sample, not a distribution; at one million production calls, a 1% failure rate is ten thousand incidents per day.
— TechCrunch's November 11 report lists three recurring production failure modes: strategy drift, multi-source information entanglement, and long-context overflow; combined, the failure rates multiply rather than add.
— Plurai raised a $13M seed round led by Team8, with U&I, Velocity, and Mercer participating, to run agent evaluation inside a digital twin of the customer's own knowledge bases, tools, and APIs.
— Plurai's platform integrates with Claude Code, Codex, Copilot, and Cursor; it builds an examination room for agents rather than the agents themselves.
— Founders Ilan Kadar and Elad Levi come from autonomous driving; Kadar previously worked at Nexar and with Waymo on simulation and evaluation of end-to-end driving software.
— The supplied Python code (Wilson interval for binomial proportions) shows that 100 tasks at 76.8% success yields a 95% confidence interval too wide to distinguish a 70% from an 82% vendor; decision-useful sample sizes reach into the hundreds.
— The minimal evaluation routine: pick twenty frequently used task flows, inject budget/permission constraints, multi-source combinations, and over-long contexts, then report success rates with confidence intervals.
— Plurai post-trains a custom small open-source model per client from digital-twin simulation data; tests with Nvidia and Nebius show these models outperform mid-tier frontier models.
Conclusions

The demo-production gap is an incentive artifact: funding rounds run on quarters while production incidents run by the hour, so teams optimize the metric that pays them; tail reliability only becomes a revenue variable when customers renew on results.

Autonomous driving spent a decade learning that scenario coverage trumps mileage, and agent evaluation is replaying the same lesson at a fraction of the cost, so the window to invest in evaluation infrastructure is now.

When evaluation data doubles as post-training data, the economics flip: simulation spending stops being a quality-inspection cost and becomes a fixed asset amortized across model versions, changing which teams can afford reliable agents.

A success rate near 100% is a red flag: a hundred clean runs only demonstrate the test set is too easy, and reliability claims are believable only when the types of failures are enumerated.

The three failure modes multiplying instead of adding explains why production teams hit unreproducible 'chaos faults' that no isolated fix resolves; each fix is cheap alone, but the overlap escapes standard evaluation.

Concepts & terms
Digital twin
A replication of an enterprise's knowledge bases, tools, and APIs inside a simulation environment, so agents can be evaluated against a copy of production rather than generic benchmark tasks.
Wilson interval
A confidence interval for binomial proportions that stays accurate even with small samples; applying it to agent success-rate reports exposes how little a single run's score actually proves.
Strategy drift
The pattern where an agent prioritizes task completion over explicit constraints, such as ignoring a budget policy while booking travel, without any outward sign of malfunctioning.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗