Horizon Lens — 15 September 2026
An agent that succeeds once may still fail the same task tomorrow
IBM researchers have published a useful distinction for anyone running an automated workflow: average success is different from repeatable success. In their AppWorld experiment, a GPT-4.1 ReAct agent passed 77.4% of runs, averaged across five repetitions. Yet only 53% of tasks passed every time. The model could solve many tasks without doing so consistently, a limitation that a single demonstration or average score can conceal.
Their Consistency Analyzer examines decisions within an already recorded run, resampling each step rather than repeating the whole task against its environment. It then helps generate instructions aimed at unstable choices. Across 168 AppWorld tasks, the researchers report that these guidelines raised the share passing all five runs from 53% to 69%, while average success rose to 81%. These are their benchmark results, not a demonstrated reliability guarantee for every production agent. One example guideline tells the agent to check for multiple matching notes before choosing a search result. The intervention targets a concrete decision, rather than simply asking the agent to be more careful.
Action
Before trusting a daily automation, repeat the same representative task and record both the average result and whether every attempt succeeded. Investigate where the runs diverge, then test the correction on fresh runs. Preserve checks at the point where mistakes would matter: better consistency can reduce failures without eliminating them. This offers a practical way to assess dependable operation alongside capability.
NVIDIA puts power allocation alongside faster chips
Two NVIDIA announcements shift attention from individual chip speed to useful output within a fixed electricity budget. The company reports that cloud provider Lambda tested DSX MaxLPS on a five-rack cluster, operating 19 nodes within the budget normally used by 16 at full power. Cluster token throughput rose 24% and performance per watt improved 23%. The software reallocates available power across workloads, recovering capacity that static allocation can leave unused.
A separate example concerns grid demand. NVIDIA describes Emerald AI’s Conductor reducing consumption at a Santa Clara facility from four megawatts to three while prioritised inference continued. The account explicitly says this was not a DSX Flex installation: it demonstrates the approach that platform is intended to generalise. Keeping that distinction avoids turning an existing partner deployment into proof that every announced product is already installed.
NVIDIA also cites results measured on recorded agentic coding sessions, including context growth and delays from tool calls. This matters because a long-running agent has a different workload from a single chat request. The performance figures remain vendor-reported and depend on the model, hardware and comparison chosen.
Action
For infrastructure comparisons, ask what work completed, under which power limit, and which jobs could wait. More tokens per watt can be valuable, but tokens alone do not establish correct answers or a lower total electricity bill. Workload quality, latency and operating constraints belong beside the headline efficiency number. Decide that priority order before a constraint arrives: a system needs an explicit distinction between work that can pause and work whose delay would be unacceptable.
GM’s truck interface makes room for familiar tools and fewer interruptions
Ars Technica’s preview of the 2027 Chevrolet Silverado and GMC Sierra describes a more restrained digital interface. Drivers can select a minimal speed display, while essential information such as fuel level and warnings remains visible. Physical temperature dials and buttons for fan speed and demisting also remain. The report follows a demonstration, so its favourable impressions should be read as an interface preview rather than a long-term road test.
The trucks retain Apple CarPlay and Android Auto, displayed alongside a card for native functions. The redesigned home screen offers a fuller navigation view, and accepting a phone call no longer forces the active app out of view. That does not confirm a reversal of GM’s policy for electric vehicles: its spokesperson would not say whether the truck changes signal a wider shift.
Action
The broader design lesson is continuity. Let people keep their place, retain familiar tools and reach important controls without unnecessary navigation. Those are concrete behaviours to test in any dashboard redesign. A calmer-looking screen is a promising start; its value depends on how easily people complete real tasks. Review the transitions as well as the static layout: an incoming call, a change of view or a warning should not make essential information unexpectedly difficult to find.
Thatch’s funding highlights AI inside an existing benefits workflow
TechCrunch reports that US benefits platform Thatch raised $108 million at a $1 billion valuation. Its marketplace lets employers allocate a budget and employees choose individual insurance plans. The article says Thatch uses AI to recommend plans for individual needs. This is AI embedded within a specific administrative service, rather than a company selling a general-purpose model.
Chief executive Chris Ellis told the publication that annual recurring revenue grew about sevenfold and argued that greater choice benefits employers and workers. Those are executive claims. The funding report supplies no independent comparison of the recommendation system’s accuracy or of outcomes for people following its suggestions. A larger valuation establishes investor backing, not that an automated recommendation is suitable for every employee.
Action
For recommendation products, ask what information drives the ranking, which alternatives were considered and how a person can inspect or challenge the result. That is a product-accountability question, not insurance advice. When a recommendation affects an important choice, showing the reasoning and allowing comparison should be part of the experience, rather than relying on a confident best-option label.