AI technology observer documenting AI evolution and industry news, sharing development practices.
The article explains how Anthropic engineers treat the limited token budget of large language models as a finite resource, detailing static configuration, runtime retrieval, and long‑task strategies such as compaction, structured notes, and sub‑agent architectures to build reliable, efficient AI agents.
Anthropic’s engineering blog outlines a systematic approach to evaluating AI agents, highlighting why agents are harder to test than traditional software, defining key concepts like tasks, trials, transcripts, and outcomes, and detailing the three grader types, evaluation timing, and practical decisions for building robust eval pipelines.
The roundup covers Anthropic’s Claude Opus 4.7 launch with improved long‑task processing and higher rate limits, Alibaba’s open‑source Qwen 3.6‑35B‑A3B sparse‑MoE model, Anthropic usage tips, OpenAI Codex’s expanded plugin suite, GLM‑5.1 tool‑call fix, Ternary Bonsai’s ternary‑weight efficiency, Tencent’s HY‑World 2.0, Sim2Reason physics learning, plus Gemini on Spot and π0.7 robot model releases.
Anthropic unveiled Claude for Word's public preview while the AI ecosystem debates Claude Mythos's looped‑language‑model architecture, faces OpenClaw compatibility hurdles, highlights new neural‑computer research, warns that memory will become the next bottleneck for agents, and questions the impact of soaring model prices on the industry.
The article analyzes Anthropic's Managed Agents design, showing how OS‑style virtualization of core components—Session, Harness, and Sandbox—creates stable interfaces that keep agent systems functional as model capabilities evolve, improve security, and boost performance.
The article analyzes Anthropic’s “Agent harness” design, showing how separating generation and evaluation into distinct agents—drawing inspiration from GANs—overcomes context‑window limits and self‑evaluation bias, enabling a three‑agent planner‑generator‑evaluator pipeline that builds a full‑stack app in six hours.
The article reviews recent AI developments, including the low‑VRAM Gemma‑4‑21B‑REAP model, Qwen3‑Coder‑Next REAP variants, Farzapedia's file‑plus‑Wiki memory system for agents, turboquant‑gpu's 5.02× KV‑cache compression, Claude API's new streaming refusal mechanism, and DeepMind AlphaEvolve's logistics savings.
The March 30 AI daily digest curates recent open‑source voice input and TypeScript libraries, new development workflows, a 30B parameter model that runs on 24 GB GPUs, and NVIDIA's PivotRL research that reduces reinforcement‑learning rollouts while matching end‑to‑end performance, all with concrete benchmarks and links.
The article explains how Prompt Caching leverages vLLM's PagedAttention and block‑level hashing to reuse KV cache across identical prefixes, dramatically cutting LLM inference latency and cost, and provides concrete engineering tips for maximizing cache hit rates.
In March 2026, attackers hijacked the official PyPI maintainer account of LiteLLM, released two malicious versions that were downloaded 46,996 times in 46 minutes, exfiltrated credentials, launched a fork‑bomb, and demonstrated how unpinned dependencies and .pth files can turn a simple package install into a full‑scale supply‑chain breach.
