The AI Agent Market Just Split Itself in Two

The AI agent landscape looks like one market. It's actually two completely different ones, and most builders haven't realised the split yet. On one side: task agents-schedulers, expense filers, inbox triagers. Clear inputs, measurable ROI, deterministic outcomes. On the other: reasoning agents-research, analysis, strategy work. Higher variance, harder to measure, expensive to make reliable. The trap is conflating them. Task agents will lose to native features from Salesforce and Microsoft once they're integrated. Reasoning agents will burn cash on reliability engineering. The founders who win are the ones who pick one market and own it completely.

Building Production AI Without Breaking the Bank

Developers building on LLMs hit the same wall: bigger context windows aren't the answer. A 2M-token window sounds great until you realise the cost, latency, and the "lost in the middle" problem-models literally pay less attention to content buried in long contexts. The real question isn't how to fit everything in; it's what actually needs to be in the prompt right now. One travel booking platform reduced their average tokens per request from 18,000 to 6,500 by combining four techniques: sliding window summarisation for recent turns, relevance-based retrieval for agent history, structured memory for facts that can't be lost, and context compression for large documents. Response latency improved from 4.2 seconds to 2.1 seconds. Better quality at roughly the same cost. The lesson: constraint forces better design.

Quantum States, Designed by AI

Researchers at the University of Tübingen trained a transformer-based language model to design quantum optics experiments. Give it a target quantum state; it outputs Python code describing the experimental setup. Out of 20 target state classes, the system produced valid construction rules for six-four matching known solutions, two completely new. The model discovered previously unknown ways to assemble optical components for specific entangled states. The experiments haven't been run in the lab yet, but they're testable blueprints. What matters here isn't the novelty of the states themselves. It's the principle: researchers can now ask an AI system to propose experimental setups without spending months exploring configurations manually. This accelerates discovery in quantum computing and quantum communication, where specially engineered quantum states are the foundation.

Beyond the technical achievement, this signals a shift in how science gets done. Rather than replacing physicists, AI is changing how they think about problems. Instead of manually assembling setups, researchers define the space of possible configurations and let algorithms explore it. The system still has clear limits-it can't guarantee perfect state fidelity and sometimes fails to find solutions. But it already contributes meaningfully to discovery, even in physical experiment design. That's a different level of collaboration than most AI tools offer.