Today at a glance

The strongest signal today is operational rather than flashy. Coding agents are becoming normal enough that their costs, permissions, security boundaries, and ownership models need real engineering discipline. Chinese-language sources were thin on deep technical posts, so the V2EX pick focuses on agent workflow behavior rather than general community chatter.

Picks

  1. Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k · Hacker News

    This comparison looks at how many tokens Claude Code and OpenCode spend before they even read the user’s prompt. That fixed overhead matters once agents run in long sessions, CI jobs, batch workflows, or high-volume team usage. The practical question is not which tool is cheaper in the abstract, but which context-loading behavior is necessary for your work and which part is avoidable tax.

  2. Old and new apps, via modern coding agents · Hacker News / Terry Tao

    Terry Tao’s write-up is useful because it deals with real apps rather than a clean demo project. Agents are often most valuable when they help inspect old systems, create scaffolding, test hypotheses, and modernize small pieces without pretending the existing codebase does not matter. That is a better adoption frame for teams with years of accumulated software.

  3. Since Chromium 148, Math.tanh is now fingerprintable to link underlying OS · Hacker News / Scrapfly

    Scrapfly shows how Math.tanh behavior in Chromium 148 can expose enough floating-point variation to help infer the underlying operating system. Browser fingerprinting keeps moving into lower-level runtime details, not just obvious headers and cookies. Privacy, anti-abuse, and bot-detection teams should treat this as another reminder that the fingerprint surface is broad and constantly shifting.

  4. Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper · Hacker News / Ploy

    Ploy reports a production agent migration to GPT-5.6 with faster runs and lower cost. The exact numbers will not generalize to every workload, but the evaluation shape is useful: measure real latency, tool calls, failure modes, and token distribution instead of relying only on benchmarks. Model migrations are starting to look like routine infrastructure work.

  5. Dicklesworthstone/destructive_command_guard · GitHub Trending

    destructive_command_guard is a Rust tool for blocking dangerous git and shell commands when agents operate in a developer environment. That is a timely problem: giving an agent terminal access without enforceable boundaries turns small mistakes into repository damage. Even if teams do not adopt this specific project, they should define hard deny rules for destructive operations.

  6. wonderwhy-er/DesktopCommanderMCP · GitHub Trending

    DesktopCommanderMCP gives Claude terminal control, file search, and diff editing through an MCP server. It is a clear example of MCP becoming the local plugin layer for agentic development. The tradeoff is equally clear: stronger desktop access requires tighter scoping, logging, review, and secret-handling rules.

  7. Directly Responsible Individuals (DRI) · Simon Willison

    Simon Willison connects the DRI concept to AI agents and argues that an agent should never be the ultimately accountable owner of a project. That sounds obvious until organizations start assigning work to agents in ticket systems and automation queues. The implementation detail is human: every agent-driven workflow still needs a named person who owns the outcome.

  8. 用 GPT Pro 20x 和 Claude Max 20x 的全部周限额跑 GPT5.6 Sol + Claude Fable 循环迭代，给富有科技做了个官网 · V2EX

    This V2EX post describes using high-limit GPT and Claude plans in a loop to build a company website. It is a messy but informative snapshot of how individual developers are combining multiple frontier models, reviews, and iterations outside formal tooling. The next hard parts are acceptance criteria, factual checking, cost control, and preventing polished output from hiding shallow work.

  9. 情報漏洩に敏感な金融機関で、Claude・Gemini・ChatGPTを導入した話 · Zenn

    This Zenn article covers the rollout of Claude, Gemini, and ChatGPT inside a financial institution that is highly sensitive to information leakage. The value is in the risk framing: approvals, data classification, user education, and governance matter as much as model choice. Enterprise AI adoption is mostly an organizational systems problem once the basic tools are capable enough.

  10. IT系上場企業の平均年収を業種別にみてみた 2026年版［後編］ · Publickey

    Publickey’s annual survey of average salaries at listed Japanese IT companies is not a deep technical post, but it is useful context for engineering leaders and developers working with Japan. Compensation patterns shape hiring, retention, outsourcing pressure, and where experienced infrastructure talent can accumulate. It is a good complement to the tool-heavy agent stories today.

Editor’s note

Today’s mix is 7 English-source items, 1 Chinese-source item, and 2 Japanese-source items. All seven requested sources were reachable; Anthropic news was skipped because there was no sufficiently fresh technical post in the last 24 hours, and weaker V2EX items were filtered out as ads, lifestyle posts, or low-signal product chatter. Dev Digest editor would start with the token overhead analysis, destructive command guard, and the financial-institution AI adoption write-up.