BestBlogs Issue #85: Harness Engineering

OpenAI's GPT-5.4 marks the first time reasoning, coding, native computer use, deep web search, and million-token context have been unified in a single model — without sacrificing performance in any individual area. The standout capability is native computer operation: the model interprets screen state and executes mouse and keyboard actions, surpassing human average on desktop benchmarks. A new tool search mechanism also cuts token usage by 47% for Agent tasks, making this a rare case where capability gains and cost efficiency arrive together.

GPT-5.3 Instant skips benchmark chasing and focuses on the user experience: fewer preachy disclaimers, sharper intent detection, and better search integration. With hallucination rates down 26.8% in connected mode, it's a meaningful step toward making ChatGPT a reliable daily tool.

Qwen3.5 releases four compact models from 0.8B to 9B under Apache 2.0, fine-tunable on consumer GPUs. The 4B impresses for multimodal and Agent tasks; the 9B rivals larger models — both well-suited for cost-efficient vertical deployment.

FireRed-OCR fine-tunes Qwen3-VL-2B into a specialized document parser via three-stage progressive training, achieving 92.94% on OmniDocBench v1.5 — the top end-to-end result. It handles formulas, tables, and handwriting at industrial quality, and is fully open-source.

A concise comparative breakdown of six leading open-weight LLMs, covering MoE design, attention mechanism tradeoffs, and post-training strategies — with practical guidance on what to actually evaluate when choosing a model. Essential reading for engineers navigating the current open-source landscape.

Pragmatic Engineer sits down with Boris Cherny, the creator of Claude Code, tracing its journey from an internal Anthropic side project to one of the fastest-growing developer tools available. Boris walks through his daily workflow — 20-30 PRs per day, 100% AI-generated, not a single line written by hand — and shares the internal debate over whether to release it at all. The conversation covers how code review is evolving in an AI-first environment, the layered security design behind Claude Code's architecture, and Boris's take on which engineering skills will matter most going forward. The printing press analogy running through the episode makes for a thought-provoking listen.

This article offers a clear framework for thinking about the developer's role in an AI-assisted workflow, identifying three positions: humans outside the loop (vibe coding), humans inside the loop (manually reviewing every artifact), and humans on the loop. The author argues the third is the right place to be — where developers shift from writing code to building and maintaining the "harness" that guides agents: the specs, quality checks, and workflow instructions that shape what agents produce. The piece goes further to describe an "agentic flywheel," where agents not only execute tasks but continuously evaluate and improve the harness itself.

The author identifies a core problem with AI-assisted coding: AI skips the design phase entirely, silently embedding every architectural decision into generated code and turning code review into an exhausting exercise in reverse-engineering. The proposed solution, "Design-First," structures the collaboration into five sequential levels — capabilities, components, interactions, contracts, implementation — with no code until the design is agreed at each step. This isn't process ceremony; it's cognitive load management that forces decisions to happen at the right level of abstraction. For developers who regularly use AI coding assistants, this article offers a practical and well-reasoned collaboration framework worth adopting.

A thoughtful piece from Alibaba's Tmall tech team: the core bottleneck in enterprise AI Coding isn't agent execution, but accurately conveying complex task goals to AI. The solution is a layered, unified expert knowledge base that reduces information entropy and drives a shift from tool-level efficiency gains to a knowledge-driven intelligent development paradigm.

A deep dive into OpenClaw's context management source code, covering its three-layer defense — preventive pruning, LLM-based compaction, and overflow recovery — plus a detailed analysis of how each operation affects Provider KV cache costs. One of the most thorough technical breakdowns of long-session context management for AI agents available in Chinese.

This piece makes a sharp argument: AI agents aren't accelerating the software development lifecycle — they're ending it. The author works through each SDLC phase systematically: requirements become a byproduct of iteration, design emerges through collaboration, tests are generated alongside code, PR review becomes a legacy ritual, and observability evolves from a passive dashboard into the feedback loop driving the entire system. The conclusion is that only two core capabilities survive: context engineering and observability. A clearly argued piece worth engaging with seriously.

OpenAI's engineering team used Codex to generate 1 million lines of code and 1,500 PRs over 5 months — with no human-written code. The article distills what made it work: structured knowledge management, rigid architectural constraints, agent-accessible observability tooling, and periodic entropy cleanup. A concise blueprint for scaling agent-driven development.

Two AI founders share their engineering workflow in the Agent era: the job has shifted from writing code to building environments for AI, a three-step flow (plan → run → validate) is now the daily norm, and judgment bandwidth — not execution speed — is the new productivity ceiling. Worth listening to for anyone curious about what AI-first engineering actually looks like in practice.

Anthropic's head of design Jenny Wen shares her firsthand observations on how the design role is transforming in the AI era. Her core argument: the traditional "discover-diverge-converge" design process is dead — not because designers chose to abandon it, but because engineers shipping at AI-assisted speed forced the change. Design work is now splitting into two modes: real-time execution support alongside engineers, and shorter-horizon vision work spanning 3 to 6 months rather than years. She also details her own workflow shift — time spent on polished mockups has dropped from 60-70% to 30-40%, with far more time now spent on paired collaboration with engineers and direct code-level work. A valuable firsthand account for any designer navigating this transition.

Zapier's VP of Product shares firsthand enterprise AI transformation practices: 800 internal AI agents, a clear distinction between adoption and transformation, and a strong argument that leadership must personally use AI tools for change to take hold. The core difference between traditional and agentic workflows? The ability to reason and dynamically reroute.

A thorough hands-on review of Gemini 3.1 Flash Image: improved text rendering accuracy, significantly better character consistency, and new support for extreme aspect ratios like 1:4 and 1:8. Comes with extensive ready-to-use prompts — useful for anyone looking to iterate quickly on design and content creation at lower cost.

The biggest mistake in AI product development is adding AI to features instead of helping users complete real jobs. This podcast presents a three-step framework — deconstruct, redesign, disrupt — and maps out four AI-native startup paths: unlocking new markets, wrapping mature tech, selling infrastructure, and transforming traditional industries. Practical and grounded with real-world examples throughout.

This three-hour conversation between Meng Yan and Li Jigang centers on a deceptively simple premise: the Industrial Revolution took physical labor, AI is taking cognitive labor, and what remains for humans is what they call "heart force" — will, intuition, and aesthetic sensibility. From there, the conversation moves through the nature of the vector world, the shift in business models from "weaving networks" to "drilling deep wells," the fork between amplifying human distinctiveness through AI versus withdrawing from thinking altogether, and the evolution of education from knowledge-pouring to spark-finding. Li Jigang's two observations — "Your feed is your fate" and "prompts have shape" — are each worth sitting with independently. Worth listening to for anyone thinking seriously about what it means to be human in the AI era, beyond the technical discussion.

Insight Partners co-founder Jerry Murdock argues autonomous agents are the real core of this AI wave — tools like Cursor are already facing obsolescence, seat-based SaaS pricing will give way to consumption models, and white-collar displacement will become an election issue within two years.

A thought experiment written from a 2028 vantage point, tracing AI's economic left-tail risks: white-collar displacement → consumer spending contraction → private credit defaults → mortgage market stress — a feedback loop with no natural brake. Not a prediction, but a rigorous risk scenario worth reading carefully for anyone thinking about AI's macroeconomic implications.