--- title: 'In the News: August 21, 2026' description: "Brightbeam AI ships an open protocol for auditing agent overrides, and Claude's post-Opus-5 verbosity keeps drawing workarounds." canonical_url: 'https://darkfactory.dev/news/2026-08-21-evening' markdown_url: 'https://darkfactory.dev/news/2026-08-21-evening.md' collection: news date_published: '2026-08-21T19:45:00-04:00' date_modified: '2026-08-21T19:45:00-04:00' --- # In the News: August 21, 2026 An open protocol for recording what happens when a human overrules an agent ships a real reference implementation this week. Two more tools join the growing pile of workarounds for Claude 5's verbose writing style, and the term this publication tracks as dark factory turns up twice, unprompted, in one Hacker News thread. ## 1. A protocol for auditing what happens when a human overrules an agent **[Collaborative Human-Agent Protocol (CHAP)](https://github.com/BrightbeamAI/chap)** · Arsalan Shahid, Gordon Suttie, Philip Black, Brightbeam AI · GitHub, technical report on arXiv 2606.09751 Brightbeam AI's CHAP 0.2 gives agent-review workflows a structured record for the moment a human edits or rejects an agent's draft. Each override becomes a hash-chained envelope carrying a diff, a rationale, and a set of controlled-vocabulary tags, so an `audit.read` call can replay the full decision chain months later instead of someone reconstructing it from chat logs and ticket comments. The protocol's most useful idea is a two-way split: a refining override means the human agreed with the agent's decision but rewrote how it was expressed, while a substituting override means the human reached a different decision entirely. Brightbeam frames these as different failure modes needing different fixes: a high refining rate points at retrieval problems, a high substituting rate points at an ambiguous policy. The repository ships a TypeScript reference implementation (`@chap/coordinator` on npm), a CLI, an override-analytics tool, and a conformance harness with 21 test vectors that check the hash-chaining and signing behavior the spec promises. CHAP is built to sit next to MCP and A2A rather than replace them. **Why it matters:** Teams running agents on review-heavy work now have a concrete shape to put override data in, instead of losing it to Slack threads and ticket comments, and the refining-versus-substituting split gives them a real diagnostic rather than a guess about whether a bad outcome traces to retrieval or to policy. The protocol is a young public draft with one reference implementation and no adoption track record yet, so treat it as a shape worth watching, not a standard to build on today. ## 2. A one-day, 294-point thread makes the case that Claude 5's voice has become a workaround problem **[Vomit](https://github.com/zachahn/vomit)** · Zach Ahn · GitHub, discussed on Hacker News since 2026-08-20 Ahn published a small command-line tool that pipes Claude 5's replies through a second, smaller local model, run via Ollama, Llama.app, or any OpenAI-compatible endpoint, to strip what he calls its token vomit before a human has to read it. The Hacker News thread carrying it collected 294 points and 286 comments in a day. The top comment, from a practitioner describing weeks spent fighting the same problem across both Claude and Codex, put it plainly: "The baked in communication style of these models is so obnoxious it's impacting my work... It's like they've been trained to produce output that's hard to read." Ahn's own README does not oversell the fix: the tool is "totally vibe-coded," slow, and its cleanup model "hallucinates a bit" because it only sees Claude's text, never the underlying files or actions. **Why it matters:** Vomit is a second inference call, and a second cost, spent undoing the first one. That a practitioner would rather build and ship that than keep prompting around the problem is the clearest signal yet that Claude's post-Opus-5 verbosity has moved from a complaint into something people spend engineering time on. ## 3. The term "dark factory" turns up unprompted, twice, in one Hacker News thread **[Building an (almost) fully self-hosted, sandboxed, agentic software factory](https://news.ycombinator.com/item?id=49390463)** · jakelsaunders94, personal blog · Hacker News, submitted 2026-08-21 Jake Saunders described, in a post drawing 63 points and 42 comments, a home-server pipeline for building software with agents rather than running it on someone else's cloud. The build drew the usual production questions: several commenters pressed on what verification looks like once a system like this reaches production, and one noted the setup was not fully self-hosted since it still leans on a Codex subscription for inference. What stood out inside the thread, rather than the post itself, was that two commenters used the term dark factory without prompting, and neither is a voice this lane already tracks. codazoda wrote they were "recently inspired by another article here to start" their own build, linking a separate series at joeldare.com documenting a minimal dark factory. mettamage put it more generally: "Quite a lot of people are automating the SDLC now and are looking at dark factory solutions." **Why it matters:** This is a small, direct data point that the term is moving past the corner of the internet that coined it and into general practitioner vocabulary, worth noting for anyone tracking how this space names itself. This run did not load Saunders' underlying blog post, so its build details are not confirmed here beyond what the thread itself describes. --- ## Also this cycle - **[nobuzz](https://github.com/adnanakil/nobuzz)** · Adnan Akil · A same-week Claude Code skill (`/debuzz`) that pipes Claude's last reply through the Gemini CLI to strip the same over-written voice Vomit targets. Drew 109 points and 89 comments on Hacker News; one commenter's fix for the underlying problem was blunter: banning comments outright in CLAUDE.md and backing the rule with a hook, because the model "forgets" style rules once its context window fills.