# X / Public Discussion Notes Source focus: Anthropic X announcement for "Harness design for long-running application development" Original X URL: https://x.com/AnthropicAI/status/2036481033621623056 Captured: 2026-06-24 ## Access Status The user asked to record the X comment section in detail. I attempted multiple public-access routes: - `x.com/AnthropicAI/status/2036481033621623056`: direct open returned no readable page lines in the browsing tool. - `publish.twitter.com/oembed`: succeeded and confirmed the original post text and date. - `lightbrd.com/AnthropicAI/status/2036481033621623056`: visible to search index, but direct fetch hit a Cloudflare verification page. - `r.jina.ai/http://r.jina.ai/http://https://x.com/...`: returned a temporary anonymous-access block for x.com. - Web search result for the X URL reports **315 replies**. So this is not a complete X comment archive. It is a verified record of: 1. the original X announcement; 2. public metadata visible through search/oEmbed; 3. comment snippets exposed by search indexes; 4. detailed public discussion from the matching Anthropic LinkedIn cross-post, because that page exposes comments without login through public web indexing. ## Original X Announcement Anthropic's X post announces the Engineering Blog article and describes it as a multi-agent harness for frontend design and long-running autonomous software engineering. Confirmed via X oEmbed: - Author: Anthropic, `@AnthropicAI` - Date: 2026-03-24 - Link: https://x.com/AnthropicAI/status/2036481033621623056 - Original article link: https://www.anthropic.com/engineering/harness-design-long-running-apps ## X Comment Signals Visible From Search Search indexing confirms a large discussion: - Search result metadata: **315 replies**. - One visible reply snippet reacts to the phrase "multi-agent harness" with surprise. - Another visible reply snippet frames the practical cost problem as multi-agent quota exhaustion: Claude Max + Claude Code, around fifteen agents, one resumed thread, and quota disappearing very quickly. Interpretation for our survey: 1. The community immediately noticed the multi-agent structure, not just the model. 2. Cost/quota exhaustion is part of the harness story, not an implementation footnote. 3. Long-running/multi-agent coding looks attractive, but its practicality depends on budget, scheduling, and when evaluator loops are actually worth the overhead. ## Matching Public LinkedIn Discussion Anthropic also posted the same article on LinkedIn: - URL: https://www.linkedin.com/posts/anthropicresearch_harness-design-for-long-running-application-activity-7442249767620861952-7qtA - Public metadata visible through indexing: about 4,001 reactions and 280 comments at capture time. This is not the X comment section, but it is a useful public companion discussion for the same Anthropic announcement. ### 1. Generator-evaluator maps to TDD and acceptance criteria One commenter connects generator/evaluator separation to TDD and mature engineering practice. The important idea is that the sprint contract resembles the conversation between engineer, product owner, and QA before implementation starts. Relevance: - This supports our idea that verifier feedback should be contract-bound. - It also suggests an experiment: compare verifier feedback with and without explicit acceptance criteria. ### 2. Long-running agents are process systems, not only code systems Several comments interpret Anthropic's harness as a move from task execution to process execution. The multi-agent structure is viewed as closer to integrating AI into real business workflows, where complexity is in management, state, and coordination, not just code. Relevance: - Our project should avoid reducing the loop to "image-to-code". - The stronger framing is sustained system work: observing, revising, validating, and maintaining coherence. ### 3. Decision authority and lineage become harder One detailed comment argues that preserving artifacts is not the same as preserving authority. Planner, generator, and evaluator can improve capability, but they also create questions: - Who owns the decision? - What authority does a decision carry? - How does that authority evolve across iterations? - How do we avoid fragmented decision lineage? Relevance: - Our verifier should not only report failures; it should also record why a fix is accepted or rejected. - This maps to intervention recording, failure attribution, and traceability in a benchmark. ### 4. Trajectory drift can happen even when local steps look consistent Another comment emphasizes that a process can remain locally consistent while gradually becoming harder to steer. Individual checkpoints may pass, but the trajectory can drift. Relevance: - Multimodal verification should include local checks and trajectory-level checks. - A UI can satisfy the current screenshot while drifting away from the intended product identity or interaction model. ### 5. Browser-based peer review is a practical frontier Some comments highlight the value of agents reviewing accessibility, browser compatibility, and UI logic before human review. Relevance: - Playwright + screenshot + DOM + accessibility tree is a natural baseline for our verifier. - We can test whether visual/behavioral QA catches issues text-only review misses. ### 6. Consistency across object definitions, behaviors, and runtime matters One comment on the retro game maker notes that creation modules can be connected from a UI perspective but still lack a continuous logic-constraint layer tying object definitions, behaviors, and execution states together. Relevance: - This is exactly the class of bug that visual-only verification may miss. - Our verifier may need to combine rendered behavior with state/model consistency checks. ### 7. Quality is structural, not only a model capability Public comments repeatedly frame quality as a structural property: a skeptical evaluator, separate roles, and explicit contracts change the system outcome more than asking one model to "be careful." Relevance: - This supports the claim that our contribution can be loop/harness design, not just a stronger VLM. ### 8. Persistent memory and context architecture are unresolved One commenter describes chunking a project into modules with agent pods, while still losing coherence, and considers persistent memory architectures such as SQL history plus semantic vector recall. Relevance: - Multimodal verification feedback needs a memory path. - If verifier findings are not persisted and re-injected, the system may rediscover the same visual/behavioral failures. ### 9. Practical automation patterns are emerging Another comment describes automating review by invoking Claude from the command line and controlling sessions, creating separate review/code sessions dynamically. Relevance: - This aligns with our desire to implement verification as a repeatable harness component. - The artifact format matters: files, logs, screenshots, and structured JSON feedback are easier to orchestrate than chat-only critique. ### 10. Dynamic skills and multi-skill problems Public discussion also raises the question of how multiple agents handle skills: whether dynamic skilling is possible and how skills should be selected or composed during a long-running process. Relevance: - Our benchmark may need to record not only verifier output, but also which tools/skills were invoked during observation and repair. ## Takeaways For Our Paper Survey 1. The comment ecosystem emphasizes the same shift we care about: from code generation to engineered production loops. 2. Community reactions are excited about generator/evaluator separation, but worried about cost, quota, memory, coherence, and authority. 3. For our project, this suggests the verifier should produce durable artifacts, not just chat feedback. 4. A strong benchmark should track not only final quality, but also cost, number of iterations, accepted/rejected findings, and whether the same failure recurs. 5. X comments could not be fully archived from public access; if a logged-in browser session is later available, this file should be expanded with a proper comment capture.