Agentic Coding

Briefing: Agentic Coding For the software engineer building with AI coding agents daily

Key Insights

• The harness is the product, not the model. Claude Code's leaked 512K-line TypeScript codebase is overwhelmingly harness—permission pipelines, 3-layer memory architecture, prompt cache optimization with 14 tracked cache-break vectors, frustration detection via regex, and multi-stage error recovery tuned from empirical data (250K wasted API calls/day before fixing). The highest-ROI skill to develop right now isn't prompt engineering but harness engineering: designing the control layer, error recovery, and memory management around whatever model you use. A personalized CLAUDE.md with your conventions and structured context outperforms chasing model upgrades.
• Claude Codeソースコード流出から学ぶハーネスパターン10選 ── 50万行のTypeScriptが明かした設計思想
• Your harness, your memory

• The Folder Is the Agent

• Unattended agents fail in specific, predictable ways—and the fix is data format, not more autonomy. Vercel's Turborepo optimization effort is the best practitioner case study available: same model, same codebase, same harness, but switching profiling output from Chrome Trace JSON to Markdown yielded "radically better" optimization suggestions and an 81–91% speedup. Unattended agents never benchmarked their own work, hyperfixated on first ideas, chased meaningless microbenchmarks, and never wrote regression tests. But over time, as the codebase improved from merged human-reviewed PRs, agents spontaneously started writing tests and creating abstractions that matched the human developer's style—the codebase itself became implicit memory.

• Making Turborepo 96% faster with agents, sandboxes, and humans

• "Dark code"—code running in production that no human ever understood—is the emerging organizational crisis, and it requires systemic response, not just better tooling. Amazon rebuilt their coding tool Kira around spec-driven development after a December outage exposed this problem. The solution has three layers: force understanding before code exists (write specs before generating), make systems self-describing (embed structural and semantic context in the codebase itself, not just docs), and implement comprehension gates (AI-assisted senior-engineer-level questions before merge). This isn't a security or quality problem—it's an organizational capability problem that reaches board-level accountability.

• I Looked At Amazon After They Fired 16,000 Engineers. Their AI Broke Everything.

• Mythos represents a step-change in coding benchmarks, but the model you should worry about configuring is the one you already have. Mythos shows SWE-bench gains of +13/+24% over Opus and Terminal Bench 2 at +16%, but it's being withheld for safety—a first. For daily work, the practical takeaway is task-matched model selection: Opus/Mythos for complex planning, Sonnet for implementation ($0.04/PR for automated review), and cross-model review where Claude writes and GPT reviews to catch biases the generating model misses. Multiple independent practitioners arrived at this cross-model pattern. MirrorCode shows Opus 4.6 reimplementing 16K lines of Go (2–17 human weeks of work), confirming these models can handle real repository-scale tasks.

• The First UNSHIPPED Model: Claude MYTHOS (Senior Engineer Breakdown)
• Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment
• GitHub Copilot CLI で分析業務を半自動化する（フェーズ分割×サブエージェント×シェルスクリプトの設計）

• Claude CodeでPRレビューを自動化する設計と実装 ── 「AIレビューだけでマージ」を実現するまで

• The Planner-Generator-Evaluator separation pattern is emerging as the practical standard for multi-agent coding, and the entry point is just adding an Evaluator. "Tuning a standalone evaluator to be skeptical is far more tractable than making a generator critical of its own work." Sub-agents in Claude Code deliberately lack main conversation context—this is a feature, not a bug, preventing context snowball. Use JSON over Markdown for inter-agent communication to prevent agents from rewriting shared files. Start by adding a single Evaluator agent to your existing workflow before building out full Planner-Generator pipelines; this gives immediate quality gains at low adoption cost.

• Claude Codeのサブエージェントを使い倒す ── Anthropic公式「計画・生成・評価」3分離パターンの実践
• Claude Code's new UI 👨‍💻, Codex Scratchpad 📝, multi-agent coordination 🤖

Emerging Patterns

• Context engineering—finding the minimal high-signal token set—is becoming as important as the code itself. The 3-layer memory pattern (CLAUDE.md for static rules, MEMORY.md for session-spanning learning, project files for decisions) appears in both Claude Code's leaked architecture and independent practitioner guides. Concrete token-saving techniques include jupytext conversion for notebooks (94% token reduction), session clearing at logical breaks, local LLM for RAG to avoid agent file exploration, model tiering (Haiku/Sonnet for formatting, Opus for planning), sandbox scoping via .claudeignore, and primitive conversation compression that strips honorifics and cushion words. A "pain count" mechanism where repeated feedback (3+ occurrences) auto-promotes items from memory to CLAUDE.md rules mirrors how operational knowledge crystallizes.
• CLAUDE.md + メモリ3階層設計で始めるClaude Code活用術 ── 初心者から中級者へのステップアップガイド
• Claude Max 20xプランでも足りないので、トークン節約のためにやったこと8選

• Claude Codeソースコード流出から学ぶハーネスパターン10選 ── 50万行のTypeScriptが明かした設計思想

• Small teams are shipping production features by treating agents as draft generators with human review gates, not autonomous coders. A 2-person engineering team used Claude Code Plan Mode before execution, parallel repo clones (one for implementation, one for plan creation), CodeRabbit for initial automated PR review with custom AI skills handling review feedback, and schema-first parallel development. The core operating principle: "AI creates drafts, humans review and decide." Non-engineers (designers, planners) doing "vibe coding" for UI micro-adjustments was acceptable as long as engineers reviewed the output. Separately, an automated PR review pipeline achieved 83% auto-merge rates with Claude Sonnet at $0.04/PR, phased in from shadow mode over months with kill switches.

• 少人数で新機能を出し切る — エンジニア2人 × AI × チームの戦い方

• Claude CodeでPRレビューを自動化する設計と実装 ── 「AIレビューだけでマージ」を実現するまで

• Terminal-first agent workflows are winning for orchestration, while IDEs add automation layers. Claude Code's architecture, job queue tools like tq for multi-session management, phased shell scripts for deterministic tasks ("routine work is shell, judgment work is Copilot"), and slash-command dispatch layers all point to the terminal as the primary orchestration surface. Meanwhile, Cursor is adding GitHub-event-triggered automations for PR review, Codex is building a Scratchpad for parallel task execution, and Anthropic is developing Coordinator Mode for Claude Code's desktop experience. The pattern: terminal for control and orchestration, IDE for interactive editing and visualization. ACP (Agent Client Protocol) is emerging as the de facto standard for connecting agents across both surfaces.

• The Folder Is the Agent
• Claude Codeのマルチセッション管理にジョブキューの概念を取り入れる
• GitHub Copilot CLI で分析業務を半自動化する（フェーズ分割×サブエージェント×シェルスクリプトの設計）
• Toad: Your AI Coding Agent Deserves a Better User Interface

Dissenting Views

• Multi-agent swarms vs. single focused agents: a direct contradiction on whether more agents help. The prevailing view from the Mythos breakdown is that multi-agent orchestration is essential and single agents unsupervised are a failure mode—cost concerns are "misplaced priorities." The dissent comes from a practitioner who built and ran multi-agent swarms for three months: "more agents didn't make me faster... when 10 of them finished simultaneously, I had 10 results to evaluate without enough context to know which ones I could trust." The resolution appears to be that orchestration complexity should match project complexity—start with an Evaluator, grow only when needed—and that the human manager's review bandwidth, not agent speed, is the actual bottleneck. This is a difference in emphasis: both sides agree multi-agent has value, but they disagree sharply on when to reach for it.
• The First UNSHIPPED Model: Claude MYTHOS (Senior Engineer Breakdown)
• The Folder Is the Agent

• Claude Codeのサブエージェントを使い倒す ── Anthropic公式「計画・生成・評価」3分離パターンの実践

• Memory ownership: open vs. proprietary is a real lock-in vector, not a theoretical concern. LangChain argues that using closed harnesses behind proprietary APIs yields control of agent memory to third parties, creating "incredible lock-in"—and that memory, not the model, is the true sticky differentiator. Claude Code's leaked source reveals the opposite investment thesis: deep proprietary memory architecture with anti-distillation mechanisms, suggesting Anthropic views memory as a competitive moat to protect. If you're building products on top of agents, this tension directly affects your architecture decisions. This is a direct contradiction on strategy, not just emphasis.

• Your harness, your memory
• Claude Codeソースコード流出から学ぶハーネスパターン10選 ── 50万行のTypeScriptが明かした設計思想

Read & Act

What to read:

• Making Turborepo 96% faster with agents, sandboxes, and humans — The most rigorous quantified case study of agentic coding in production. The discovery that LLM-friendly data formats dramatically improve output quality is immediately applicable, and the iterative workflow (profile → review → implement → benchmark → PR) is adoptable this week.

• Claude Codeソースコード流出から学ぶハーネスパターン10選 ── 50万行のTypeScriptが明かした設計思想 — The deepest technical analysis of how a production agentic coding tool actually works. The 3-layer memory architecture, prompt cache optimization strategy, and empirical tuning data provide a blueprint for anyone building or configuring agent harnesses. Read this to understand what Claude Code is doing behind the scenes.

• I Looked At Amazon After They Fired 16,000 Engineers. Their AI Broke Everything. — Names the "dark code" problem and provides the most structured organizational response. If you're a tech lead or above, the 3-layer solution (spec before code, self-describing systems, comprehension gates) addresses a blind spot that most practitioner discussions miss entirely.

• The Folder Is the Agent — The honest counterweight to over-engineering multi-agent systems. Context drift, agent stalls, and "you can't vibe orchestrate" are lessons from months of running a dispatch layer across dozens of agents. Read this before building anything more complex than a single project folder.

• Claude Codeのサブエージェントを使い倒す ── Anthropic公式「計画・生成・評価」3分離パターンの実践 — The most practical implementation guide for Planner-Generator-Evaluator, including why sub-agents lack main context, JSON vs. Markdown for communication, binary evaluation criteria, and a cost-benefit framework for when multi-agent is worth the 3–5x expense.

What to do:

• Reformat one data source your agent reads and measure the difference. The Turborepo case proved that switching profiling output from JSON to Markdown—same model, same data—produced "radically better" suggestions. Pick your most-used agent workflow, identify the ugliest input it processes, and make it human-readable. If it's hard for you to scan, it's hard for the agent. Run the same task before and after and compare output quality.

• Add an Evaluator sub-agent to your highest-traffic workflow this week. Define it as a separate Markdown file in .claude/agents/ with binary pass/fail criteria focused on observable facts (does the function handle errors? does the test cover the edge case?), not subjective quality. Use JSON for inter-agent communication to prevent rewrite accidents. This is the lowest-cost entry point to multi-agent patterns and delivers immediate review quality gains.

• Audit your CLAUDE.md and memory files against the 200-line/25KB budget. Strip vague instructions ("write clean code") and replace with verifiable rules ("functions under 50 lines," "always read the file before writing"). Move domain knowledge into .claude/rules/ with path-specific frontmatter so it loads only when relevant. Implement session clearing at logical breaks. Track which rules the agent actually follows—if it routinely ignores one, the rule is either too vague or too buried.