Voice Agents Are Finally Real. Your Architecture Isn't.The launch of GPT-Realtime-2 landed quietly for a supposedly major release. No benchmark fireworks, no "state of the art" claims. Just three streaming audio models and a developer guide about reasoning effort, preambles, and long-session state manage...May 11, 2026·4 min read
Correctness Before Corrections: What vLLM's RL Migration Teaches Us About Agent ReliabilityMay 8, 2026·3 min read
The Services Layer Is Eating Your Agent StackThe Services Layer Is Eating Your Agent Stack Anthropic just raised a billion and a half dollars for a services joint venture. OpenAI spun up a ten billion dollar deployment company. If you're still thinking about agents as API calls wrapped in Pytho...May 6, 2026·4 min read
Tool Use Is the Easy Part. Knowing When Not to Call One Is the Work.Tool Use Is the Easy Part. Knowing When Not to Call One Is the Work. Most agent failures don't look like failures. They look like success at the wrong thing. I've watched production agents call a search tool fifteen times in a row, each query slightl...May 4, 2026·5 min read
Evaluation Is Now Your Infrastructure BottleneckEvaluation Is Now Your Infrastructure Bottleneck We spent years optimizing training. Distributed strategies, memory-efficient optimizers, flash attention — the whole stack got compressed and accelerated until the marginal cost of training another epo...May 3, 2026·3 min read
The AI pricing floor just dropped. Your infrastructure bets need to change.Most production teams are optimizing for the wrong thing. They benchmark latency against GPT-4, compare token costs to Claude, and build caching layers around assumptions that were true six months ago. Then DeepSeek V4 drops, and the math changes ove...Apr 29, 2026·4 min read
Vibe Coding Won't Fix Your Agent's Memory ProblemVibe Coding Won't Fix Your Agent's Memory Problem Everyone's launching an agent course now. Google's 5-day intensive. Kaggle competitions. Tweet threads promising production-ready agents in fifty lines of code. The message is clear: building agents h...Apr 28, 2026·4 min read
Your Agents Aren't Ready for Cheap Long ContextYour Agents Aren't Ready for Cheap Long Context DeepSeek V4 shipped with a million-token context window that costs less to run than Claude's 200K. The technical achievement isn't the parameter count or the benchmark scores—it's the Compressed Sparse ...Apr 25, 2026·3 min read