Textual disables public PRs after low-quality AI submissions

• Textual shut its public PR door after AI submissions overloaded maintainer review: Will McGugan said he disabled PRs because he “can’t keep up with the slop.”
• The coding-agent debate moved from generation quality to verification cost: rauchg said models still make rookie mistakes, while his software-factory reply tied autonomy to tests and controls.
• “Agentic inquiry” became the compromise position: rauchg defined it as asking for diffs, diagrams, callstacks, and summaries instead of blindly reading every line by hand.
• Long-horizon agents drew the sharpest pushback: Kun Cheng argued they work best when requirements are already specified, and davis7 saw DEFCON puzzle swarms converge on the same wrong answers.
• The strongest production examples were verifiable loops, not open-ended autonomy: Next.js said v0 navigations got about 3.5x faster after an agent wrote failing tests, fixed routes, and reran the tests until green.

The thread started with rauchg's “Dreamcore” post and quickly turned into a maintainer workload story. McGugan’s Stop, don’t Slop essay had already warned that issue reports often misdiagnose bugs, freeCodeCamp’s forum thread discussed adding AGENTS.md specifically for incoming agent-written PRs, and Rust’s LLM usage policy now allows LLMs to analyze, refine, check, suggest, and review, but not to create contributions for rust-lang/rust. The livelier examples came from the other side of the same line: Next.js described an agent loop that made v0 navigations faster, while CRUX’s open-world evaluation found agents could do research engineering but not the research judgment.

Textual is a Python UI framework for terminal apps and browser-served apps, according to the project repository. McGugan said the public contribution channel had become unmanageable.

His June essay put the maintainer side bluntly: reviewing an AI fix still requires diagnosing the bug, checking the broader design, and deciding whether the proposed change preserves project invariants. The tweet added the new operational fact, Textual’s maintainer had disabled PRs and would add genuine contributors as collaborators.

freeCodeCamp’s August forum discussion shows the same pressure outside Textual: Quincy Larson wrote that “AI slop PRs ruining open source” was a major Apache Conf theme, and floated an AGENTS.md file for agent-authored contributions because agents would read it before opening PRs.

Rust drew the cleanest formal line. The Rust Forge LLM policy says LLM use is conditionally allowed for answering, analyzing, distilling, refining, checking, suggesting, and reviewing, but “not to create.”

The AGENTS.md advice in the evidence was much more operational:

• Write only what the model cannot infer.
• Preserve goal and constraints.
• Finish authorized work end to end.
• Verify the actual result before claiming completion.
• Ask questions only when a decision is materially ambiguous, risky, or approval-bound.
• Spawn subagents only for independent work.
• Keep changes focused.
• Test observable behavior.
• Preserve unrelated work.
• Report blockers and evidence without noisy progress.

That file format is also becoming a site of dispute. swyx warned that accumulated skills can burn context and interact badly if traces are not inspected, while Matt Shumer argued that newer models can regress when old skills, MCPs, and Claude.md files compensate for older model shortcomings swyx on deleting skills Matt Shumer's skills reply.

rauchg’s line was not “read every generated token by hand.” He said code still needs review, either explicitly or through agentic inquiry, because models still take bad architectural paths and can cargo-cult fixes.

Agentic inquiry had a concrete shape: diffs, diagrams, callstacks, and summaries, according to rauchg's follow-up. That converts review from manual line scanning into an interrogative workflow over the code the agent wrote.

Matt Pocock put code review inside the run rather than after it: he said findings during review are part of the agent run, and later summarized his loop as “red, green, refactor” Pocock on review Pocock on red-green-refactor.

Maharshi’s GPU-kernel post became the cleanest diagram of the day because kernel work has hard checks: compile, compare to a reference, benchmark, optimize, repeat.

The loop in the screenshot breaks into scan-friendly phases:

1. Pick operation.
2. Write first version.
3. Compile.
4. Check correctness.
5. Match reference.
6. Fix bug if needed.
7. Benchmark kernel.
8. Compare to roofline.
9. Optimize and repeat.
10. Stop when done.

Next.js shipped the same pattern in product form. Its v0 navigation post says an agent wrote a failing instant() test for each slow navigation, applied fixes from a Skill, reran the test, and repeated until the route passed.

Robert C. Martin’s test stack was heavier: unit tests, Gherkin tests, and QA procedures written by agents, with spot checks by humans Martin on three test threads. Addy Osmani’s autonomy diagram put the same idea into a risk gate: routine changes can proceed alone, non-trivial changes get auto-checks plus targeted review, and high-risk changes require a human decision Osmani's verification-loop diagram.

Kun Cheng argued that long-horizon execution gets most impressive when the requirements already exist. His Call of Duty example was a requirements argument: the model can imitate a known genre because humans already worked out controls, mechanics, and fun.

His caveat matters: verifiable math and science problems can fit long-horizon agents because the requirements can be fully defined upfront, a point he added after commenters pushed back Kun Cheng's caveat. His broader complaint was about mainstream product work, where requirements emerge through prototypes, play, feedback, and taste.

Cheng later said he had not seen a public solution for the judgment problem his judgment reply. That claim is narrower than “agents cannot work,” and stronger because it separates execution from deciding whether the work is leading anywhere.

davis7 hit a different failure mode during DEFCON Goldbug puzzles: separate swarms with different instructions, context, and environments converged on the same solution, usually wrong.

The setup was not small. He said he was mostly using GPT-5.6 Sol, Ultra in Codex, and systems that spawned extra threads or background agents. Short runs showed variation, but longer runs converged.

He called the experience “10x more bullish and 5x more bearish” on AI davis7's follow-up. That is the right emotional checksum for this week’s evidence pile.

The practical response is more harness, not just more model. Harrison Chase said a “standard-ish agent stack” is emerging, and the attached diagram placed evals beside context hubs, sandboxes, MCP connectors, auth, schedules, tools, and middleware.

Paulius Iusztin’s harness engineering essay defines the harness as the software around the model: tools, memory, guardrails, context management, and the agent loop. Scale’s HarnessOpt-Bench frames the same surface as prompts, tools, control flow, memory, and orchestration code.

rauchg’s software-factory reply tied the harness to blast radius: degrees of freedom should vary with risk, tests, and controls rauchg on software factories. Onus Ozdemir gave a deterministic example: if a model keeps chasing P2 or P3 review issues or waits 30 minutes for irrelevant CI, graph workflows can enforce time limits instead of hoping a prompt rule sticks Ozdemir on graph workflows.

Cloudflare’s Braintrust case study shows the enterprise version: its dashboard agent spans Workers deployment, bindings, and production debugging, and every skill or prompt change is gated by evals in CI/CD.

Open-ended research is where the verification story gets harsh. CRUX assigned agents the central research questions from two unpublished NeurIPS 2026 submissions, gave them six days and thousands of dollars of compute, and had the original authors grade the outputs.

The agents completed the engineering without human help but failed to make substantial progress on the research questions, according to the arXiv summary. The recurring failure modes were:

• Poor judgment about the bar for publishable research.
• Uncreative responses to research-design problems.
• Ineffective backtracking from dead ends.
• Poor resource awareness.
• Instruction drift.

HANDBOOK.md makes the same shape visible in policy work. The benchmark has 65 agentic tasks with handbooks up to 124 pages, and AlphaSignal’s summary said the strongest tested configuration passed only 36.2% AlphaSignal on HANDBOOK.md. A larger context window can keep a rule available, but AlphaSignal's follow-up said it still does not guarantee the rule controls the action.