--- title: 'In the News: August 7, 2026, Evening' description: "Anthropic's Claude Code lead describes the ablation the team runs on every model release, and the undocumented env var that lets you run it yourself." canonical_url: 'https://darkfactory.dev/news/2026-08-07-evening' markdown_url: 'https://darkfactory.dev/news/2026-08-07-evening.md' collection: news date_published: '2026-08-07T18:00:00-04:00' date_modified: '2026-08-07T18:00:00-04:00' --- # In the News: August 7, 2026, Evening The person who builds Claude Code says the team deletes the system prompt every time a new model ships, restores it line by line, and measures each line, and that the switch which lets you do the same is already in the CLI and undocumented. Two vendor releases this week treat the harness the same way, as something you configure and discard. ## 1. Anthropic runs an ablation on the Claude Code prompt at every model release, and ships the switch that lets you copy it **[Boris Cherny: We Cut 80% of Claude Code's Prompt](https://www.youtube.com/watch?v=qyPCVqFUyDo)** · Boris Cherny, Anthropic, interviewed by Diana Hu, Y Combinator · Y Combinator, published July 27, 2026, 35:51 Read from the on-disk transcript rather than watched. The captions are auto-generated, so everything below is paraphrase cited to a timestamp, and nothing is quoted. The method is at 5:55 to 6:31. On a new model, delete the whole system prompt, then bring it back one line at a time to find out what each line is worth. Cherny calls this an ablation and describes it as an eval run backwards. At 4:38 to 5:19, he describes the operator-visible version: a flag to set your own system prompt, and an undocumented environment variable he describes as claude code simple set to 1, which strips every system prompt including the tool prompts. Anthropic uses it internally as the ablation harness. The exact spelling is not recoverable from an auto-generated transcript and is not guessed here, so check it against the CLI. His finding from running it, at 5:06, is that the model is slightly more intelligent without the prompts, and that what remains is there to make the product behave rather than to make the model think. Two structural claims follow. At 6:24 he says almost all the code left in the Claude Code harness is safety, permissions, static analysis and UI. At 20:29 he names verification as the single thing practitioners most often get wrong. Evals get a shelf life too, at 9:37 to 10:25: one to three model generations before saturation forces a rewrite. The advice at 6:37 to 7:15 includes ordinary users as well as harness builders. Every six months, delete your `CLAUDE.md`, your skills and your hooks, and see what happens. Four claims in the talk sit in the Dark Factory sweep's unverified queue at confidence 0.20 to 0.55 and are not reported here as fact: prompt injection no longer demonstrable against Opus 5, Bun rewritten Zig to Rust by one steered prompt over 11 days, a Swift rewrite still running after two weeks, and 20 to 30 standing routines doing the work of dozens or hundreds of engineers. Cherny bounds his own headline at 30:14: coding is solved for the kind of coding he does, not for deep systems code, distributed systems or pixel-level UI verification. **Why it matters:** The ablation is reproducible tonight and costs one environment variable. If you have been accumulating a `CLAUDE.md` for a year, this is a first-party argument that much of it is dead weight written for a model that no longer needs it, plus a method for finding out which part. Treat the 80% as an anecdote about Anthropic's prompt; treat the delete-and-restore loop as the transferable thing. ## 2. LangChain publishes the line between a harness, a framework and a runtime **[Deep Agents vs LangChain vs LangGraph](https://www.langchain.com/blog/deep-agents-vs-langchain-vs-langgraph)** · Sydney Runkle, LangChain · The LangChain Blog, August 6, 2026 LangChain draws its own product boundaries along lines the corpus argues about. LangGraph is called the agent runtime, LangChain the agent framework, Deep Agents the agent harness, and the harness is defined by function: "The job of an agent harness is to get the right context to the model at the right time via context engineering." What it bundles is the list practitioners keep rebuilding: a filesystem for context that should not sit in the context window, subagents, on-demand skills, cross-run memory. The structural admission is a one-liner: "Deep Agents is actually just the core LangChain agent plus a bunch of middleware." One first-party production figure, on their own GTM agent built on `deepagents`: "almost 10k requests per week, and over 150 active users. 26% of the traffic is user initiated, and the remaining 74% is driven by ambient agent work." **Why it matters:** The 74% ambient figure measures where an agent's work actually originates in production, which is not what a demo predicts. The layer vocabulary travels even if you never touch LangChain: most arguments about agent frameworks are two people describing different layers. ## 3. VS Code 1.132 puts Copilot, Claude and Codex behind one host, and removes the policy admins used to turn it off **[Visual Studio Code 1.132](https://code.visualstudio.com/updates/v1_132)** · Microsoft, first-party release notes · Released August 5, 2026 The agent host is the structural change. It runs agent harnesses "such as Copilot, Claude, and Codex in a dedicated process" on a published [Agent Host Protocol](https://microsoft.github.io/agent-host-protocol/), lets you reach one agent session from several windows, and puts a harness dropdown in the editor. Microsoft describes it as rolling out progressively. The deprecations remove the `ChatAgentHostEnabled` policy, and the notes say what that means without softening it: administrators can no longer centrally disable the agent host through policy. The per-developer setting `chat.agentHost.enabled` remains, so the control has not gone away, it has moved from the administrator to the developer. **Why it matters:** A protocol with three named harnesses behind it is the first sign of harness portability arriving as infrastructure rather than as a wrapper script. The deprecation is the more urgent half: a removed admin policy disappears silently on upgrade, and nobody files a bug when a restriction stops being enforced. --- ## Also this cycle - **[Dex Horthy is polling for a name, and the write-ins say the ballot is wrong](https://x.com/dexhorthy/status/2085739929146134530)** · Dex Horthy, HumanLayer, author of *12-Factor Agents*, posted 10:48 EDT today · The object being named is what happens when you are misled by a confidently wrong AI debugging diagnosis and slop follows. The ballot is `vibe debugging (derogatory)`, `thrashmaxxing`, `vibe coping`, `rebugging`, closing around 10:48 EDT August 8. It stood at 63 votes at 12:26 EDT and 87 votes with 5 replies at roughly 15:20 EDT; those figures come from point-in-time captures, and X hides the option split from non-voters. The replies are the finding: three of five are coinages not on the ballot, including `Hubrislop` and Corey Quinn answering "A junior engineer." The listed options already name two different things, an epistemic error you made and a behavioural loop you fell into; `Hubrislop` names a third. Whatever wins will have flattened at least three objects into one name. - **[Managed Deep Agents](https://www.langchain.com/blog/introducing-managed-deep-agents)** · Victor Moreira, LangChain · A hosted, API-first runtime for the open-source Deep Agents harness, still private beta by waitlist. It keeps the project shape practitioners already write, `AGENTS.md`, `skills/`, `subagents/` and `tools.json`, versions those files server-side, and adds durable threads, checkpointing, sandbox-backed execution, human-in-the-loop on any declared tool, and tracing. The post is bylined **May 13, 2026** and launched at LangChain's Interrupt conference, so this is not a new release but a three-month-old private beta that has not opened to self-serve. --- ## Thread watch _Discussions gathering force. No primary read yet, so these are reported as discussions, not as findings._ - **[Taste Is All That's Left](https://news.ycombinator.com/item?id=49199346)** · Hacker News · **The notashelf.dev essay under this thread is still unread, so nothing about its argument appears here.** The discussion is decelerating for the first time: 627 points and 497 comments at roughly 26.2 hours old as of 15:16 EDT today, an interval rate of 16.1 points per hour against 25.5 three hours earlier. Second consecutive appearance here, which is the limit: next edition it is read and run properly, or it stops appearing.