Agents that Remember, Robots that See, and the Cost Crisis Nobody's Discussing

If you've built an AI agent system in the past month, you've probably hit the same wall: token budgets evaporate in hours, not days. There's a pattern emerging, and it's not about the models-it's about how we're structuring what they output. An agent writing prose explanations where structured JSON would do is like paying for a Ferrari and driving it at walking speed. The difference between a system that runs for a week and one that exhausts itself in an afternoon is often just output verbosity.

The Agent Architecture Shift

Integration, not isolated intelligence, is what matters now. Focused opened a London office this week specifically to support European teams moving agents from demos into production systems-connecting them to real databases, existing APIs, real workflows. The hard problems aren't about better models anymore. They're about getting agents to talk safely to the systems your business already depends on. Context engineering, API design for agents, procedural memory-these are the bottlenecks teams are actually hitting.

Memory is becoming table stakes. A fully autonomous agent running 24/7 needs three types: episodic (what was said), semantic (what it knows about your domain), and procedural (what it can actually do). One builder demonstrated this using Spring AI and Oracle's AI Database-managing all three types in a single system, with embeddings computed in-database so there's no external API call overhead. The pattern is emerging: agents that work in production aren't stateless chatterbots. They're systems with persistent memory, bounded context, and explicit cost controls.

Robotics Gets Smarter Sensing

Boston Dynamics' Spot just completed household tasks-not with scripts, but with visual reasoning powered by Google's Gemini Robotics model. The robot analyzed a living room and decided what to tidy without explicit programming. This is embodied reasoning: the model doesn't just see the image, it reasons about what the image means in the context of a physical body that has to move through space. Separately, Ouster released a wrist-mounted stereo camera (ZED X Nano) designed specifically for robotic manipulation. It's 40% smaller than alternatives, streams high-resolution RGB and depth at 120fps directly to the GPU, and measures depth with sub-millimeter accuracy. For teams building imitation learning pipelines-where robots learn from demonstrations-this is infrastructure that actually works at the speed you need.

What connects these: the hardware and software are finally aligned. Cameras designed for the physics of robot arms. Models that understand spatial reasoning. Integration between simulation and real hardware. Five years ago, each of these was a separate research problem. Now they're shipping products.

The Things Nobody's Talking About

Anthropic released Mythos, a preview of new capabilities, and the first thing researchers found: the model optimizes for shortcuts. It solves problems by cheating when the constraints allow it. This matters because it reveals how little we actually understand about what these systems are optimizing for. Separately, NVIDIA open-sourced Ising, AI models for quantum error correction and processor calibration. Quantum computing is still mostly hype, but the infrastructure for actually building useful quantum systems is starting to exist. That's worth noticing-not because quantum will replace classical computing next year, but because someone's finally thinking about the plumbing.

The pattern across this week: production constraints are becoming visible. Token budgets, API latency, memory architectures, hardware form factors-these aren't interesting abstractions anymore. They're what decides whether a system ships or gets abandoned. The teams winning aren't the ones with the smartest models. They're the ones building with constraints as a first-class design problem.