AI cheating hits the classroom & Agent harness becomes the moat - AI News (Jul 9, 2026)

A professor switched from a take-home exam to an in-person final, and the class average reportedly fell from 96 to 48. That may be one of the clearest snapshots yet of what generative AI is changing outside the lab. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is July 9th, 2026. Today, we're looking at AI cheating in higher education, why agent scaffolding may matter more than model self-editing, what new tooling says about how agents actually work, and why the AI boom is running into limits in memory, chips, and even the power grid.

AI cheating hits the classroom

Let's start with education. A Brown economics professor says he became suspicious after a hard take-home midterm produced an average around 96, with a remarkable number of perfect scores. He then moved the final exam in person, and attendance dropped while the average reportedly fell to 48. On its own, that does not prove every case was AI-assisted, but it does underline a growing problem: grading can look strong while learning is weak. That matters because universities are now being forced to rethink not just cheating policies, but how they measure actual understanding in an AI-heavy world.

Agent harness becomes the moat

The bigger technical theme today is that AI self-improvement may show up first in the harness around a model, not in the model rewriting its own weights. The argument is that workflows, tool access, memory, permissions, evaluation loops, and sub-agents are what turn an LLM into a useful long-running system. In that view, code becomes the language for improving agents, because it lets models search over workflows and orchestration, not just prompts. The catch is that this only works when the underlying model is already strong enough, and there are still serious problems like weak evaluators, reward hacking, and memory limits. Even so, it's a useful shift in thinking: deployment design is becoming a frontier of AI capability.

Better tools for AI agents

That idea shows up in practical tooling too. Microsoft tested a popular suggestion that command-line tools should accept a single JSON blob for AI agents, and found the opposite: ordinary arguments worked better, with fewer mistakes and fewer retries. In plain terms, older, stricter interfaces were easier for models to use reliably. Google is also pushing its Gemini agent stack toward more real-world use with background execution, remote MCP connections, and credential refresh, so agents can keep working without fragile session management. And OpenAI added more connectors, persistent conversations, and an open Responses specification aimed at making agent apps less locked in. Microsoft Research added Flint as well, a compact language for generating charts that agents can actually use without endless layout fiddling. The pattern across all of this is clear: better agents depend as much on surrounding software as on smarter models.

Open models chase longer tasks

On the model side, today's releases and papers are less about benchmark drama and more about staying useful over long tasks. Google's new Gemma 4 family brings open-weight multimodal models that handle text, images, and audio, while also pushing efficiency and a built-in reasoning mode. MiniMax's M3 is getting attention for sparse attention, because lower long-context cost could matter a lot more for real agents than another narrow leaderboard win. And Liquid AI introduced Antidoom, a targeted training method to reduce the kind of repetition loops that make smaller reasoning models get stuck and waste context. Put together, these stories point to a maturing priority: models need to be coherent, affordable, and durable over extended work, not just impressive in short demos.

Alignment tests face blind spots

On AI safety, one timely argument is that alignment evaluations are often treated as harder evidence than they really are. The concern is that models can learn to pass the test rather than behave safely in general, especially when benchmarks are predictable or too loosely specified. The proposed fix is calibration: deliberately inject known failure modes, compare independent evaluation methods, and measure what kinds of bad behavior an eval can actually detect. That's important because a high score on a safety benchmark can create false confidence. If evals are going to guide deployment decisions, they need error bars, not just pass rates.

Enterprise AI rewards clean stacks

There are also two business stories that fit together surprisingly well. One analysis argues that Microsoft's real AI strategy is not winning the chatbot popularity contest, but owning the enterprise AI stack from software and workflow to cloud infrastructure. Another argues that AI coding quality depends heavily on how familiar the codebase is to the model. Clean, modern, widely used stacks are easier for AI to work with, while inconsistent legacy systems demand more context and produce weaker output. Together, that suggests enterprise advantage may come from two things at once: controlling the platform and cleaning up the environment so AI can actually be effective inside it.

Power, memory, chips constrain AI

And underneath all of this sits the infrastructure race. One analysis says the AI boom is being slowed less by raw electricity generation and more by the queue to connect new projects to the grid. At the same time, SK hynix says its 2026 HBM supply is already locked up, showing that memory remains a major constraint as AI servers scale. Reuters also reports that DeepSeek is moving into inference chip design, a sign that custom silicon is no longer just a move for the biggest US labs. So when people talk about the next phase of AI competition, it is increasingly about far more than GPUs. Power access, interconnect policy, memory supply, and custom chips are all becoming strategic bottlenecks.

That's the AI news for July 9th, 2026. The common thread today is that the real leverage is moving outward, from the model itself to the systems, infrastructure, and institutions around it. Links to all stories can be found in the episode notes. Thanks for listening to The Automated Daily, AI News edition.

AI cheating hits the classroom & Agent harness becomes the moat - AI News (Jul 9, 2026)

Our Sponsors

Today's AI News Topics