Today at a glance

The strongest picks today are less about model demos and more about operating the tools around models. Browser automation for agents, edit-tool constraints, MCP’s connection model, and token resale abuse all point to the same thing: AI systems need product-grade boundaries. The non-AI picks are also good engineering reads, especially proof automation and Decker’s lightweight creation model.

Picks

  1. We have proof automation now · Hacker News

    Adam Langley’s post uses zstd and Lean to show proof automation moving closer to real systems work. This matters because the useful version of formal methods is not a separate academic ritual; it is something that can sit near normal development, review, and CI. Security-sensitive libraries, compression code, crypto, and compilers are the obvious places to watch.

  2. Decker, a platform that builds on the legacy of HyperCard and classic macOS · Hacker News

    Decker is a small creative platform inspired by HyperCard and classic Mac software. Its appeal is that it lets people make interactive documents, tools, and demos without the weight of a full modern app stack. As AI makes code generation cheaper, lightweight environments for testing and sharing ideas become more valuable, not less.

  3. An Inside Look at the Relay Market Powering Token Resellers and Fraud · Simon Willison

    Simon Willison highlights an investigation into discounted LLM token resale, proxy pools, abused trials, exposed bots, and payment fraud. The engineering lesson is direct: any public AI endpoint without hard limits, authentication, and monitoring can become part of someone else’s arbitrage loop. Budget caps and abuse controls are not nice-to-have features for LLM apps.

  4. citrolabs/ego-lite · GitHub Trending

    ego-lite is a browser built for AI agents running web automation, including workflows that need logged-in state. That is a real pain point for agents: web tasks often fail at session handling, browser isolation, and not interrupting the user. The hard questions are credential boundaries, audit logs, replayability, and what happens when an agent clicks the wrong thing.

  5. CoreBunch/Instatic · GitHub Trending

    Instatic positions itself as an open-source alternative to Webflow, Framer, and WordPress, with self-hosted visual CMS workflows and static output. It is interesting because AI-assisted site generation still leaves the maintenance problem behind. Permissions, plugins, content models, database state, and deployment control matter long after the first page is generated.

  6. 不跳就锁——用心率广播实现人走锁屏 · V2EX

    This V2EX thread describes using heart-rate broadcasts to lock a machine when the user walks away. It is a small hack, but a good one: wearable signals, desktop automation, and endpoint security are a useful combination. The practical concerns are false positives, privacy, manual override, and making sure the automation fails in a predictable way.

  7. Code Agent 的 edit 工具有没有在生成阶段就防幻觉的方案？ · V2EX

    This discussion asks whether an AI code agent’s edit tool can reduce hallucinations during generation rather than relying only on post-edit checks. That is exactly where agent UX needs to improve. Structured patches, target-file constraints, context validation, small diffs, and rollback paths are better primitives than letting a model rewrite large files freely.

  8. Opus5が思考が浅いように感じる問題への対策 · Zenn

    This Zenn post looks at cases where Opus 5 can feel like it is reasoning too shallowly and suggests mitigations. The broader lesson is that model upgrades can invalidate old prompt habits, especially workflows built around long rule lists or implicit constraints. Teams should version prompts, context templates, and evaluation examples the way they version other production assets.

  9. MCP仕様が明日アップデート、7月28日版MCPからはステートレスな接続が正式仕様に · Publickey

    Publickey reports that MCP’s July 28 spec update will make stateless connections part of the official specification, with GitHub’s MCP server already announcing support. MCP is moving from basic tool connection into a more serious operational layer for agents. Reconnection behavior, auth, session state, and client compatibility will matter for any internal agent platform.

  10. Introducing Claude Opus 5 · Anthropic

    Anthropic’s Opus 5 launch is still worth reading because the surrounding discussion is shifting from benchmarks to safety, prompt injection, and long-running task behavior. The launch post emphasizes reasoning, coding, and proactive work, but teams should pair it with the system card and pricing details. A model release is not an adoption decision; it is the start of an evaluation.

Editor’s note

Today’s 10 picks are distributed as HN 2, GitHub Trending 2, Simon Willison 1, V2EX 2, Zenn 1, Publickey 1, and Anthropic 1. All requested sources were reachable; V2EX items that were ads, relationship threads, or non-engineering chatter were skipped. Dev Digest editor’s top reads are the LLM token resale investigation, MCP’s stateless connection update, and the Code Agent edit-tool discussion.