BestBlogs Issue #84: Orchestration

Anthropic released Claude Sonnet 4.6 during the Spring Festival, with broad upgrades across coding, computer use, long-context reasoning, and agentic planning, along with a 1-million-token context window. Notably, it outperforms the previous flagship Opus 4.5 in user preference tests 59% of the time, while keeping the same price point as Sonnet 4.5 — making frontier-level agentic capability genuinely accessible at scale.

Google's Gemini 3.1 Pro jumps from 31% to 77% on the ARC-AGI-2 reasoning benchmark and introduces a three-tier thinking mode that lets developers tune compute intensity per task without juggling multiple models. What makes this more than a benchmark story is the price: performance surged while costs held steady, with API spend coming in at under half of Claude Opus 4.6 — a calculated counterattack that put Google back at the top of the intelligence index rankings.

This hands-on review tests Nano Banana 2 across Chinese text rendering, complex UI generation, manga storyboarding, and subject consistency, offering a more grounded take than the official blog: speed and quality improvements are less dramatic than advertised, and some edge cases actually regress from the previous model. But with API pricing cut in half and a genuinely lower barrier to entry, the model makes high-quality image generation practically accessible in a way its predecessors didn't quite manage.

GLM-5, the latest open-source flagship from Zhipu AI, is built around agentic engineering as a first-class design goal, combining an asynchronous RL framework with sparse attention to dramatically cut inference costs while boosting performance. It achieves state-of-the-art results among open models on benchmarks like SWE-bench and BrowseComp, and demonstrates end-to-end software engineering capabilities that rival top closed-source systems in real-world tasks.

MiniMax M2.5 reaches top-tier performance across coding, tool use, and office productivity, but the more significant story is two structural bets: first, baking complex task decomposition directly into the model through a native Agent RL framework, achieving 37% faster completion while reducing token consumption; second, driving continuous Agent operation costs below one dollar per hour, turning the vision of economically unconstrained Agent deployment from aspiration into practical reality.

Seedance 2.0 represents a genuine leap in AI video generation — it doesn't just produce footage, it understands cinematic thinking, autonomously handling shot composition, emotional pacing, and scene transitions. The author walks through a range of creative applications, from dramatic short films and fan recreations to product ads and real-world video editing, while candidly reflecting on what it feels like to watch 18 months of hard-built workflows become obsolete almost overnight.

Claude Code creator Boris Cherny traces the full arc from accidental prototype to paradigm-shifting tool, anchored by one core philosophy: build for the model six months from now, not today. He shares that he no longer writes a single line of code by hand since Opus 4.5, and digs into multi-agent system design, the right way to use ClaudeMD, and why the most valuable skill for engineers in a rapidly evolving model landscape is no longer technical expertise — it's beginner's mind and first-principles thinking.

Writing an AI Skill is fundamentally different from writing documentation for humans. This article uses skill-creator as a case study to reveal the core principle: every line must earn its place in the context window. Defining what the AI should not do proves more precise than describing what it should, and the key to quality lies in knowing when to lock down behavior with scripts versus when to guide with natural language.

Sherwin Wu, OpenAI's engineering lead, offers a rare inside look at how AI tools are reshaping software teams: 95% of engineers use Codex daily, top performers ship 70% more PRs, and the engineer's role is shifting from writing code to orchestrating AI agents. He also breaks down why so many enterprise AI deployments fail to deliver ROI, and explores the underappreciated second and third-order effects of the emerging one-person billion-dollar company era.

As LLM scaling yields diminishing returns, memory engineering is emerging as the next core infrastructure challenge. This talk systematically covers three key mechanisms — hierarchical memory modeling, proactive scheduling, and mind-map-style information organization — with a central insight that cuts through current RAG limitations: instead of reactive retrieval that blocks inference, shift memory preparation into the idle gaps between user interactions, so the right context is ready before the query arrives.

Real-world deployment of Vibe Coding tools at Alibaba's scale surfaces challenges that benchmarks rarely capture: inconsistent code quality, longer debugging cycles, security vulnerabilities in AI-generated code, and runaway token costs. The team's practical responses — templating successful task patterns to lift completion rates, replacing closed-source models with domestic alternatives, and treating Agents themselves as composable tools — offer a grounded blueprint for anyone building or evaluating the next generation of AI coding products.

PingCAP CTO Huang Dongxu draws on hands-on experience to trace AI Coding's leap from assistant tool to autonomous agent, arguing that everything is converging toward a "Coding Agent" paradigm. He explains why context engineering has become the key performance differentiator, introduces his concept of "Box" isolated environments for safer multi-agent collaboration, and reflects on how engineers can stay relevant as the technical barrier to coding disappears.

OpenAI's CTO of Applications and Codex engineering lead offer a rare inside look at how AI coding has evolved from assistance to true delegation inside OpenAI — engineers now shut their laptops, go to meetings, and return to find the work done. What's equally striking is the chain of shifting bottlenecks: solving code generation surfaces code review, then deployment, then understanding user needs. The emerging competitive edge for engineers is no longer writing code, but product intuition and the ability to move fluidly across layers of abstraction.

A product manager with no coding background built a fully functional personal AI Agent in a single afternoon using Claude Code. The article shares six practical Vibe Coding techniques, with a central insight that cuts through the hype: AI has eliminated the technical barrier to building, but the ability to define what to build and why remains irreplaceable — and for non-technical builders, that product intuition is now their greatest competitive advantage.

Anthropic's 2026 Agentic Coding Trends Report maps out a fundamental shift in software development — from writing code to orchestrating agents that write code. Spanning eight trends from multi-agent collaboration to programming democratization, the report's sharpest insight is that AI amplifies existing engineering judgment rather than replacing it. Skills like system design, task decomposition, and quality evaluation become more valuable as implementation gets delegated to agents.

Google Chief AI Scientist Jeff Dean traces the full arc of AI infrastructure evolution through his own experience: from the 2001 decision to load Google's entire search index into memory, to the co-design philosophy behind TPUs, the logic of distillation and sparse models, and his two forward-looking bets — personalized models that can attend to everything you've ever seen, and specialized hardware driving ultra-low latency that will fundamentally reshape how humans and AI systems collaborate.

Claude Code's lead Boris Cherny traces the product's journey from an internal post that got two likes to driving 4% of GitHub's global code commits. His sharpest take: coding as a problem is largely solved, and the next frontier is AI that proactively surfaces ideas rather than just executing them. In this transition, the rarest skill will not be writing code but the generalist capacity to define what is worth building across multiple domains.

Naval reframes AI-era competitiveness through several sharp lenses: vibe coding turns taste and judgment directly into productivity, but only those who understand the underlying architecture can patch the leaks when AI makes mistakes; human agency — driven by desire and survival instinct — remains the one moat AI cannot replicate; and the only antidote to AI anxiety is opening the hood to understand how it works, then acting on that understanding.

In this interview, "father of UML" Grady Booch offers a historical antidote to today's AI anxiety: software engineering has survived multiple existential crises, and each one gave way to a new golden age. He pushes back sharply on claims that AI will fully automate software engineering, arguing that the real work of software engineering — navigating tradeoffs across competing forces — cannot be automated. What we are witnessing is simply another rise in abstraction level, and the engineers who thrive will be those who develop systems thinking and the judgment to manage complexity at scale.

Wanlater's year-end review covers the full 2025 AI landscape across seven dimensions: the rise of reasoning models sparked by DeepSeek R1, the emergence of Agent as a paradigm, the talent and organizational battles among ByteDance, Alibaba, and Tencent, the investment surge and real-world limits of embodied intelligence, and finally, the human dimension — how people navigate skill devaluation and the search for meaning in an era of accelerating automation.