Gemini 2.5 Flash Thinking API 실전 가이드 — thinking_budget을 직접 실험해서 배운 것들
Gemini 2.5 Flash의 Thinking Budget을 Budget=0/1024/8000 세 가지로 단순 작업·수학 추론·코드 리뷰에 직접 실험했다. 단순 작업은 5배 느려지고, 수학 문제는 오히려 출력 토큰을 줄여준다. 작업 유형별 최적 설정 프레임워크를 공유한다.
jangwook.net
Kim Jangwook
Personal technical notes on AI agents, automation, developer tools, and the process of building software.
Latest Notes
Recent writing by language
The root page stays intentionally small. Choose a language, then read the full archive and related posts there.
한국어
AI 에이전트, 자동화, 개발 도구, 소프트웨어 제작 과정을 한국어로 기록합니다.
Claude Managed Agents의 Dreaming, Outcomes, Orchestration — 에이전트가 자면서 스스로 나아지는 구조
2026년 5월 Code with Claude SF에서 발표된 Dreaming, Outcomes, Multiagent Orchestration 기술 구조를 분석한다. Harvey 6배 수치의 맥락, LLM-as-judge 제품화, 자기 개선 루프의 거버넌스 리스크까지.
Cloudflare Agents Week 2026 분석 — AI 에이전트가 클라우드 인프라를 직접 프로비저닝한다
Cloudflare Agents Week 2026 전체 발표 심층 분석 — Sandboxes GA, Artifacts, Dynamic Workers, 에이전트가 직접 인프라를 자율 프로비저닝하는 기능까지. @cloudflare/agents SDK 로컬 실험 결과 포함.
English
Personal notes on AI agents, automation, developer tools, and building software.
Gemini 2.5 Flash Thinking API: What I Learned from Running Budget=0, 1024, and 8000
I ran Gemini 2.5 Flash's thinking_budget at 0, 1024, and 8000 across simple tasks, math reasoning, and code review. Simple tasks got 5x slower with no gain. Math reasoning actually reduced output tokens. Here's the task-by-task decision framework.
Claude Managed Agents' Dreaming, Outcomes, and Orchestration — How Agents Self-Improve While You Sleep
A technical analysis of Dreaming, Outcomes, and Multiagent Orchestration announced at Code with Claude SF 2026. Breaking down Harvey's 6x claim, the productization of LLM-as-judge, and the governance risks of self-improving agents.
Cloudflare Agents Week 2026 — When AI Agents Become Cloud Customers
Cloudflare Agents Week 2026: Sandboxes GA, Artifacts, Dynamic Workers, and AI agents that autonomously provision cloud infrastructure. Deep-dive into 20+ announcements with @cloudflare/agents SDK experiment results.
日本語
AIエージェント、自動化、開発ツール、ソフトウェア開発の記録です。
Gemini 2.5 Flash Thinking API 実践ガイド — thinking_budgetを直接実験して学んだこと
Gemini 2.5 FlashのThinking BudgetをBudget=0/1024/8000の3パターンで、単純タスク・数学推論・コードレビューに実験した。単純タスクは5倍遅くなり、数学問題では逆に出力トークンを削減する。タスク別最適設定フレームワークを公開する。
Claude Managed AgentsのDreaming・Outcomes・Orchestration — エージェントが眠りながら自己改善する仕組みの分析
2026年5月Code with Claude SFで発表されたDreaming・Outcomes・Multiagent Orchestrationの技術構造を分析する。Harvey 6倍数値の文脈、LLM-as-judge製品化、自己改善ループのガバナンスリスクまで。
Cloudflare Agents Week 2026分析 — AIエージェントがクラウドインフラを自律的にプロビジョニングする
Cloudflareが4月のAgents Week行事で20件以上の発表を行った。Sandboxes GA、Artifacts、Dynamic Workers、さらにエージェントが直接Cloudflareアカウントを作成してドメインを購入する機能まで。@cloudflare/agents SDKをローカルで実際に動かした結果をまとめた。
简体中文
记录 AI 代理、自动化、开发工具和软件构建过程。
Gemini 2.5 Flash Thinking API 实战指南 — 亲测 thinking_budget=0/1024/8000 的真实结论
我用 Budget=0/1024/8000 三种配置，在简单任务、数学推理、代码审查三类场景下亲测了 Gemini 2.5 Flash Thinking API。简单任务速度慢 5 倍毫无收益，数学推理反而减少了输出 token。分享按任务类型选择最优 budget 的决策框架。
Claude Managed Agents的Dreaming、Outcomes与Orchestration — 代理在睡眠中自我改进的结构分析
深度解析2026年5月Code with Claude SF大会发布的Dreaming、Outcomes和多代理编排技术架构。剖析Harvey 6倍数据的背景、LLM-as-judge的产品化，以及自我改进循环的治理风险。
Cloudflare Agents Week 2026分析 — AI智能体成为云基础设施的自主运营者
Cloudflare Agents Week 2026全面解析 — Sandboxes GA、Artifacts、Dynamic Workers，以及AI智能体自主提供基础设施的功能。20多项公告深度分析，附@cloudflare/agents SDK本地实验结果。