Clanker Rage, MCP Goes Stateless & DHH Flips to GPT-5.5

Agentic Coding & Agent Harnesses

Armin Ronacher's clanker rant goes viral. A one-line vent — "Expected a change to be 10 lines of code. Clanker made a 300 lines diff. In moments like this I want to throw the damn thing against the wall" — picked up 13.8K views and 348 likes in a few hours. When pressed, Ronacher named the culprit: GPT-5.5. No, it wasn't tests bloating the diff; the model just rewrote everything in sight. The replies turned into one of the better meta-threads of the week, with @guilhermeotina nailing it: "the gap between 'correct code' and 'minimal correct edit' is where the real taste lives. llms are great at the first, terrible at the second." And @noxfield405: "scope control is the missing primitive. the agent pays nothing to write 300 lines. you pay everything to read them." Several recommended explicit anti-bloat instructions in the prompt ("minimize diff size, backtrack elegantly"). Thread.

swyx ships Kakuna. A direct response to the same problem: a skill suite that only hardens existing code — checklists, subagent parallelism, opinionated takes on how AI engineers should design apps for joint human/agent access. The framing: "instead of dark factory, go mullet factory — party in front (ship unique lovable features), dark in the back (timeless production principles)." You /plan with it, then let /goal run for a day, and it returns the same functionality with the boring stuff done plus an audit of its own work. This is the productized version of yesterday's experiment where swyx let an agent run for 16 hours, churn out 103 commits, and end up with "exactly the same app but instead of fragile mvp it now looks like a codebase i can actually build on." Repo. Kakuna announcement.

Matt Pocock floats TEST-SEAMS.md. A third doc layer alongside CONTEXT.md and ADRs — an explicit list of agreed test seams in the app. The motivation: "Agents simply cannot be trusted to make good decisions about what to test, and at what seam. For every small change, they extract out only what they've built into a testable function and test that. It leads to a patchwork nightmare of tests that break as soon as the implementation changes." Asked how to keep it from going stale, Pocock said merge it into every planning skill that touches module shape (/to-prd, /improve-codebase-architecture). 42 replies, 27K views — clearly a nerve. @scratchdotmd offered the dark take: "it seems to work best when the agent is unaware of the existence of tests." Thread.

Uncle Bob endorses the tuning tax. Bob Martin — quoted approvingly by Pocock — said 30–40% of his agent productivity goes into tuning agents and hardening output, "and I don't consider that a waste." His tell: the code isn't necessarily better, but the surrounding tests are vastly better than anything he produced manually, even with disciplined TDD. Pocock's gloss: "Uncle Bob gets it." 412 likes, 97K views. Thread.

Codex /goal graduates from experiment. OpenAI shipped GA for goal mode in Codex (app/IDE/CLI), and notably you can now pause and steer mid-goal rather than waiting for it to finish flailing. swyx tested it: "lol ok you can pause and adjust mid goal now." In a parallel demo, @reach_vb handed Codex a screenshot of swyx's challenge — code a ~10M-parameter transformer in JAX/Flax/Optax in free Colab and train it on addition — and Codex drove a logged-in Colab session via Chrome, ran the T4 job, and audited the result with subagents. Final: 10,652,557 params, ~19 min train, 99/100 exact accuracy. /goal announcement. Colab demo.

Model & Tool Releases

DHH: GPT-5.5 has overtaken Opus 4.7 for complex agent work. "For complicated agent work, it's amazing how much GPT5.5 has improved. I found 5.2 to be very far behind Opus. Now using Opus 4.7 after 5.5 feels like a big step backwards. Gotta love this level of competition! Strong comeback for OpenAI." 3,151 likes, 225K views — clearly resonated. Several replies echoed the pattern: Codex/GPT for building, Opus still better for content. Thread. Worth pairing with Ronacher's GPT-5.5 complaint above — same model, opposite verdicts, depending on whether you prize raw capability or restraint.

Simon Willison releases Datasette Agent (alpha). A conversational AI assistant for Datasette that answers questions about SQLite databases and is plugin-extensible for extra tools. Asked how it compares to "Claude Code + SQLite," Simon was candid: "Claude Code + SQLite is hard to beat one-on-one!" — Datasette Agent's edge is running as a browser web app with multiple-model support and the existing Datasette plugin ecosystem. Announcement · Blog post · Live demo.

MiniMax M3 confirmed open weights. Ryan Lee from the MiniMax team confirmed M3 will ship open-weight, "soon." LLMJunky's reaction: "awesome. this guy cooks." Quote.

Protocol & Ecosystem

MCP 2026-07-28 release candidate is out and the protocol is now stateless. David Soria Parra: no handshake, no session id, any request can hit any server instance. Plus extensions as first-class citizens (MCP Apps, Tasks), auth hardening, and a formal deprecation policy "so we don't have to do this again." 1,397 likes, 561K views — easily the biggest infra-story of the day. @heygurisingh summarized the point well: "Making MCP stateless is one of those changes that sounds boring until you realize it massively simplifies scaling, reliability, and orchestration for real production agents." The cost: sampling is gone. Ronacher's reaction: "Finally. I asked for this back in March last year!" Thread · Blog.

GitHub starts shipping native PR contribution caps. Camilla Moraes teased a "PR cap for outside contributors and an allowlist for trusted ones," aimed at the post-Shai-Hulud-era flood of low-quality PRs. Peter Steipete: "We used bots so far to enforce a 10PR per 'person' limit. Great to see GitHub shipping that natively!" Preview.

Mitsuhiko on Rust appeal post-curl CVEs. "Now that even pillars of the C ecosystem like curl have legitimate CVEs filed against them like there's no tomorrow, I wonder if Rust is about to look a lot more appealing even to the critics." A reply asked the obvious counterpoint: "if the models are so good at finding security issues, maybe the value of memory safe languages is diminished?" — which Ronacher has not yet rebutted publicly. Thread.

Industry & Markets

SpaceX's IPO filing is wild. LLMJunky's breakdown — Q1 numbers:

  • Space: lost $662M
  • Connectivity (mostly Starlink): made $1.1B
  • AI: lost $2.5B

The only profitable line is Starlink, but the valuation pitch is AI — SpaceX claims a $28.5T TAM, 93% of which is AI. Over 60% of Q1 capex went to AI infrastructure. Highlights from the AI segment:

  • A $1.25B/month GPU compute deal with Anthropic that can be cancelled in fewer than three months
  • A teased $60B Cursor deal
  • Grok improvements

On governance: Musk's Class B shares get 10 votes each vs. 1 for Class A; a non-voting Class C exists for acquisitions. The filing says Musk will control the company. Thread.

This is the same Anthropic-SpaceX deal that surfaced as a rumor on May 21 — now confirmed in S-1 form, with the cancellable-on-3-months clause being the obvious load-bearing detail for anyone modelling Anthropic's compute risk.

Theo on dev platforms. A side note from a day of off-topic posting: "Crazy how the web is still the only platform that doesn't suck to develop for." No further argument given, but the implicit comparison is iOS/Android/Windows. Post.