The Reliability Gap in Agentic AI
This week's three papers converge on a single uncomfortable question for any executive betting on agentic AI: how much of what looks like model capability is actually the scaffolding around it? Each paper answers differently, but the direction is the same — the scaffolding is doing more of the work than most vendors are letting on, and the scaffolding is where your leverage lives.
Fragility You Can't See in a Demo
The first paper, from a team at Salesforce AI Research, tears into the 'self-improving agent' narrative that has dominated 2026 pitch decks. When they simply ran the same agents multiple times and shuffled the task order, headline gains collapsed. The default task orderings used in prior papers were acting as a hidden curriculum — a sequence carefully arranged so early wins seeded later ones. Reorder those tasks and the agent's 'learning' looks a lot more like luck. For a CIO evaluating an agent platform, this is a procurement red flag: if the vendor's benchmark used a single run in a fixed order, you have no idea what you'll actually get in production.
The Two-Sided Consent Problem
The second paper, based on surveys of nearly 5,500 dating-app users, is the first serious empirical look at what happens when both sides of an interaction can deploy autonomous agents. The finding is stark: people are roughly three times more willing to send agent-mediated messages than to receive them. This asymmetry is not a dating quirk — it is the shape of every emerging agent-to-agent market. Whether it's AI SDRs contacting AI gatekeepers, or procurement bots negotiating with sales bots, the party who benefits from delegation is not the party who has to tolerate it. Design for that asymmetry or your rollout stalls.
Rubrics Beat Intelligence
The third paper is the most immediately actionable. When grading open-ended answers, the identity of the answer explained 95.6% of score variance; the identity of the judge model explained 0.2%. Small cheap models graded as reliably as frontier ones — as long as they had an explicit rubric with an official answer. Strip the rubric and reliability collapses. The design pattern the authors propose, 'any-to-bench,' is the operating manual: spend once on a frontier model to extract the rubric at ingestion, then run all repeated scoring on the cheapest model that can read. For any business running LLM-based evaluation, classification, or QA at volume, this is a direct path to 10-100x cost reduction with no measurable quality loss.
What These Three Have in Common
All three papers say the same thing from different angles: the model is not the product. The evaluation protocol, the consent mechanics, and the rubric are the product. Vendors selling you 'the model' are selling you the least differentiated piece.
On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification
What they did: Ye and colleagues at Salesforce AI Research took two published memory-based self-improving agents — the kind that maintain a textual memory bank and supposedly get better as they see more tasks — and did something the original papers didn't: they ran them multiple times, and they shuffled the task order. Then they inspected the memory the agents built up to understand where the fragility came from.
Why it matters: The self-improving agent is a load-bearing concept in the current wave of agentic AI marketing. If it holds, you get compounding value from every deployment. If it doesn't, you get an expensive stochastic system that looks competent in demos and unpredictable in production. This paper's finding is closer to the second reality. Agent gains were highly dependent on the specific task ordering used in the original benchmark, and the noise from a single agent run compounds when you stack a learning loop on top. Adding rubrics and environment feedback helped, but did not close the gap — meaning there are still uncharacterized failure modes.
What it means for business: Change your procurement checklist. Any agent platform claim of 'learns on the job' should be validated across at least five independent runs with shuffled task orders on your own workload. If the vendor can't or won't do that, assume the demo number is an upper bound you will not hit. And invest in the boring layer everyone underweights: explicit task specifications, clear rubrics, and human-oversight interfaces. That layer is where reliability actually lives.
Delegation Asymmetry in Agentic Recommender Systems
What they did: Leshchikova and colleagues fielded two large surveys (N=2,894 and N=2,617) on a major dating platform in two languages, measuring separately how willing users are to (a) deploy an autonomous agent to converse on their behalf and (b) receive agent-mediated messages from others. They used a latent-variable measurement model and validated it against a held-out sample.
Why it matters: The two willingnesses are highly correlated (0.92) but statistically distinct, and the asymmetry is huge. The threshold for deploying your own agent is -0.38 on their scale; the threshold for engaging with someone else's is +0.32, and for full engagement, +1.39. Mean deployment propensity is roughly three times mean engagement propensity. Under random pairing, only 4-13% of interactions would combine a willing sender with a willing receiver, with pronounced gender-directional imbalance.
What it means for business: The economics of agent-to-agent markets are not symmetric. A reciprocity requirement (agents only talk if both sides opted in) cuts volume by more than half. A receptivity-aware routing rule (only send agent messages to people who welcome them) tripled per-contact engagement in this study. For any customer-facing agent — sales outreach, scheduling, support triage — build in a disclosure and opt-in layer from day one. The market will not tolerate one-sided delegation, and the platforms that recognize this asymmetry first will build the trust that lets everyone else scale.
Grading Needs a Rubric, Not Intelligence
What they did: Jhen-Ke Lin ran a controlled experiment on LLM-based grading: six cost-efficient model configurations across two families at three reasoning-effort levels, each grading 24 open-ended exam questions three times, for 3,456 grades. Then two ablations — remove the rubric criteria, remove the official answer — to isolate what the rubric was actually contributing.
Why it matters: The variance decomposition is striking. Answer identity explains 95.6% of score variance. Judge identity explains 0.2%. Raising the judge's reasoning effort moves scores by at most 0.006 of full marks. Frontier judges are no more reliable as a panel than the cheap ones. But strip the official answer out of the rubric and reliability collapses from ICC 0.888 to 0.628, and suddenly judge intelligence starts to matter again. The rubric — specifically the official answer within it — is what decouples grading quality from model size.
What it means for business: If you are running any evaluation, classification, moderation, or QA workload on frontier models at volume, you are almost certainly overspending. Follow the 'any-to-bench' pattern: use the frontier model once, at ingestion, to author the rubric and the reference answer. Then route all repeated scoring to the cheapest model that can follow instructions. Cost drops by an order of magnitude with no measurable quality loss — assuming, and this is the whole game, that you actually invest in writing the rubric well. That is the work the frontier model can't do for you.
Key Takeaways
• Self-improving agents are far noisier than single-run benchmarks suggest — always evaluate across multiple runs and shuffled task orders before trusting a production deployment.
• The 'implicit curriculum' in demo task ordering is often a hidden prerequisite for agent success; scramble the order and the gains frequently collapse.
• Underspecified tasks and environments amplify agent fragility — investing in explicit rubrics, feedback signals, and oversight interfaces materially reduces failure modes.
• Consent to agent-mediated interaction is asymmetric: people will happily deploy their own agent but resist engaging with someone else's, which will constrain agent-to-agent commerce.
• Receptivity-aware routing (matching only where the recipient actually welcomes agent contact) tripled per-contact engagement in the dating study — a design pattern that generalizes to sales, outreach, and support.
• For grading and evaluation workloads, a well-written rubric matters vastly more than model size: 95.6% of score variance came from the answer, only 0.2% from the judge model.
• Any repetitive evaluation pipeline should follow the 'any-to-bench' pattern — spend once on a frontier model to extract the rubric, then run all repeat scoring on cheap models.