写作
英文文章
筛选文章 全部文章
编辑精选
精选文章，先从这里开始
先看方法、架构与工程实践，再按主题继续阅读。
Should an agent keep going or start a fresh attempt?
Use When Agents Slow Down, Reflexion, Hyperband, and agent evaluation to compare continued revision with independent attempts under reproducible budgets.
阅读要点比较连续修改与独立尝试，按相同预算设计评测，并检查最终结果能否通过验收。
阅读文章What Tau solves by splitting an AI agent into three layers
Tau separates model adapters, the agent core, and the application so tools and frontends can work together without depending on each other.
阅读要点分清模型适配、Agent 核心与应用层，检查工具和界面之间的依赖。
阅读文章Connecting an agent loop to a frontend with events
Follow tool execution and repeated model calls, separate streamed deltas from final messages, and explore frontend state with a model-free Python demo.
阅读要点用 Python 示例观察工具调用、错误和结束事件，区分界面更新与会话记录。
阅读文章专题、教程与项目实践
第 1 页 · 12 / 70-
EmDash 1.0 separates plugin catalogs from publishing EmDash 1.0 separates author identity, signed releases, plugin discovery, and runtime permissions. Installers verify records independently.
-
Agents API browser auth keeps passwords out of model input How Agents API separates site approval, login credentials, and screenshots, and why network access and site approval cannot replace action-level confirmation.
-
Instant agent steering and recoverable tool calls A close reading of Codex instant_interrupt: ending model output, continuing background tools, and preserving state when users change requirements mid-run.
-
Cloudflare redesigns its CLI for agents Cloudflare launches cf in open beta, using Wrangler usage data to explain default JSON, command search, and typed configuration.
-
OpenShell enforces agent permissions outside the workload NVIDIA OpenShell places sandbox, network, credential, and policy controls outside the agent workload, with a Supervisor enforcing permission boundaries.
-
Keep review evidence separate from compacted agent context Independent Codex Guardian history shows why generation context and review evidence need separate records, with provenance through resume and rollback.
-
Cloudflare data puts agent traffic on the website budget Cloudflare and Fastly network data reveal repeated crawling, origin costs, access-policy questions, and the returns websites get from machine traffic.
-
MentraOS runs Miniapps on phones for multiple glasses MentraOS runs Miniapps in a phone runtime, sharing a connection to manage input, output, hardware capabilities, and permissions across smart glasses.
-
Why old CLAUDE.md rules can hinder a model upgrade Use prompt audits, instruction layers, and migration evals to retain real requirements, reassess model workarounds, and enforce hard limits in code.
-
Company Brain separates team memory permissions Company Brain scopes team-agent memory by room and user permissions, with personal authorization, traceable sources, and separate tool-write access.
-
Microsoft connects AI-built tools to app delivery After Copilot Code builds an internal tool, Managed Runtime handles Git, dev environments, Entra ID, connectors, previews, deployment, and governance.
-
Why agent network permissions need ongoing checks Codex network policy changes explain revocable permits for redirects, long responses, and WebSockets, and why tool and network permissions need separate layers.