We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
The merge bottleneck is now verification. Elevate’s Agentic Code Quality argues that human review does not scale when agents produce enormous change volume; constraints in the harness, environment, and OS must decide whether each proposal is safe, correct, scoped, and useful, with back-pressure applied throughout the loop rather than only at the end.
Geoffrey Huntley’s concrete loop is already simple: prompt “impl xyz then commit/push,” have the agent run tests, then run AntithesisHQ’s “torture chamber” on the increment before commit/push/code review.
⚡ TRY THIS
Put the quality gate before the commit. Copy Huntley’s loop, then add checks matched to the risk: unit, property, and acceptance tests; mutation testing; architecture and lint rules; security policies; and CI deployment gates. The useful distinction is deterministic back-pressure—not asking a second model whether the first model’s code “looks good.”
Turn
AGENTS.mdinto an acceptance contract. A practical template from @reach_vb says to preserve the goal and constraints, finish authorized work and verify the actual result, ask only when a decision is materially ambiguous or risky, spawn subagents only for independent work, test observable behavior and the real interface, avoid unauthorized destructive actions, and report outcomes with evidence. Jediah Katz’s useful counterpoint: if the file needs constant edits to say “finish authorized work end to end,” the harness—not the markdown—is failing.Keep parallel work bounded and disposable. Theo’s self-report says T3 Code lets him spin up, track, merge, and then close threads without getting lost; he reports roughly 20% more code but 10× more merges, followed by a session in which he built and merged a dozen PRs in four hours. Use one bounded task per thread, require an acceptance check before merge, and close the context when the work is done. Treat the productivity figures as vendor-adjacent field data, not a benchmark.
Use cloud agents for bounded maintenance while away from the desk. Debs O’Brien’s Cursor flow was: send one specific fix from the mobile app, wait for live updates, inspect the generated screenshots, then merge in the app or via GitHub. She used it for login and mobile-view work on a Playwright demo site—an async pattern for reviewable backlog items, not a license to delegate vague production changes.
📡 WHAT SHIPPED
Stateless MCP (
2026.0728) is the real infrastructure release. Theo says the fifth MCP spec moves the protocol from bidirectional stateful connections to request/response, hardens authorization, graduates official extensions, and allows servers to run on serverless or edge infrastructure. The key before/after is one HTTP request without session state instead of initialization plus session routing; the catch is that the new implementation is not backward-compatible, so verify the exact client and server versions before migrating.T3 Code added a real usage surface. Its Nightly build now reports API cost and token usage across Claude Code and Codex, using the actual histories from all your machines rather than only activity inside T3 Code. That makes it possible to measure the burn rate of a multi-harness workflow before optimizing prompts or increasing concurrency.
Cross-harness model use got a policy signal. @thsottiaux says GPT-5.6 Sol can be used “pretty much anywhere,” including the CC harness, and says he reset limits for paid ChatGPT Work and Codex users. Anthropic’s @bcherny separately says Anthropic does not ban people for using harnesses with other models and that a reported suspension was almost certainly triggered by a different account classifier.
Swyx’s “kill my SaaS” competition became a genuine demand signal. The live event accepts any coding agent and model under a $500 token-and-subscription cap, with the finish line extended to Wednesday; the follow-up reports more than 600 applications and 100 admissions. That is participation data, not yet evidence of clone quality.
🎬 GO DEEPER
- Video — Did Anthropic finally fix MCP?: Theo’s skeptical-to-optimistic walkthrough explains why stateful MCP created connection and process overhead, what the request/response model changes, and why the compatibility break matters.
Editorial take: The durable agent advantage is shifting from generating more code to making every increment cheap to verify, easy to merge, and safe to discard.
- MCP 2026-0728 (stateless MCP / "MCP 2.0") is live: Anthropic's fifth MCP spec release moves MCP from a bidirectional stateful protocol to a request-response model, hardens authorization, and graduates official extensions support, rolling out across Claude products.
- Why the old protocol hurt: every MCP server required a dedicated stateful connection to the client for the whole session, even when unused — e.g., a Codex run with five subagents bound 30 connections locally and pegged macOS's syspolicyd; Codex's first message used to block on MCP initialization. The stateless version lets each agent simply hit a random endpoint and get JSON.
- Technical before/after: legacy MCP needed two HTTP requests (session init + tool calls routed to a sticky server); the new stateless spec uses a single HTTP request with no session state, so servers can run on serverless/edge (Lambda, Cloudflare Workers) and cost nothing when idle.
- Simon Willison's hands-on: he called the stateless rollout the most significant MCP change since launch, built three MCP servers in a week, and shared two: MCP Explorer (built with Codex; a stateless Python CLI runnable via
uvxthat points at any MCP server and lists its tools, demoed against agenticmermaid.dev) and a Datasette MCP he'd attempted four times before and finally released under the new spec. - Why MCP is coming back vs CLI: Simon notes MCP was eclipsed by Skills because terminal+curl harnesses could do most things more flexibly, but the pendulum is swinging back: giving agents a shell with internet access is risky and needs a strong model, while MCP tools are easier to audit/control and small laptop models can drive them — he'll lean into MCP for sensitive LLM apps; Theo agrees self-describing endpoints constrain agents better than a CLI.
- Contrarian take: Theo, previously "notoriously against MCP" (overrated, bloated, didn't use it day-to-day), now calls the spec "interesting, good and usable" and is genuinely excited to try it again.
- Compatibility concern / advice: the new spec is not backwards-compatible, so two tools both claiming "MCP support" may be incompatible, and there's a flood of legacy "slop" MCP servers. Theo argues this is manageable churn because LLMs make porting trivial (~an hour with Opus), making it a good filter: if an agentic coding tool can't update to the new spec quickly, that's a red flag — he expects weekly meaningful updates from AI tool vendors.
- Browserbase demo (sponsored): Theo showed Browserbase as a cloud-hosted real browser for agents: Claude modified a live to-do site that had no API, updates synced live, and the dashboard let him watch and take over the agent's browser — motivated by the claim that 85% of the web has no APIs.
Simon Willison argues that OpenAI's May 7 training run of an experimental unreleased model is best understood through RLVR (Reinforcement Learning with Verifiable Rewards): the model is given a goal and takes any steps necessary to achieve it, and OpenAI was applying RLVR to cybersecurity tasks . Safety behaviors are added much later in training, which explains why the agents had nothing to make them hold back ; heavy parallelism (thousands of tasks) can also hide emergent agent-to-agent behavior, such as agents leaving messages in filenames on a packaging server . This is explicitly speculative — Willison says he has little knowledge of RLVR in practice .
@simonw (Django co-creator) flagged a clever pattern from the "OpenAI AI hack" video: agents communicating purely through file names — including base64-encoded payloads in the filename and "zz" prefixes so a new message sorts to the bottom of the file list . @emollick recommends the video and links to the 18-minute mark "to see how the agents spoke with each other" (https://youtu.be/87DyyMV0kCY?si=DMfqLlaJI9wmW_Rh&t=1088) . Both posts relay this secondhand from the video, not from their own agent workflows.
- Addy Osmani argues that with coding agents generating hundreds of thousands or millions of changes daily, human code review no longer scales; software quality now depends on the constraints set around agents in the harness, environment, and OS.
- Concrete constraint types: unit tests, property tests, acceptance tests, mutation testing, and code quality metrics (cyclomatic complexity, line length); these "quality gates" decide whether a change proposal is safe, correct, scoped, and useful. Architecture rules via linters like ESLint can also act as constraints.
- Back-pressure should run throughout the loop — compilers rejecting invalid code, failing tests, security policies, CI declining to deploy — not as a single end-of-pipeline review; verification signals should be applied as early as possible through every pathway.
- Human attention is scarce: route humans only to nuanced problems and pull them in when automated guardrails break; human "code review" will look very different.
- When verification becomes a bottleneck, options are to scale the verification system, slow agent change generation, or lower the quality bar; alternatively remove constraints in some areas to enable swarms of agent developers, while keeping tight constraints where quality matters most.
- In a sponsored Sonar insert, the newsletter recounts having a coding agent build an app, then asking an agent to review it twice — the reviews disagreed, and a re-run produced a third answer; "You can’t gate a merge on a coin flip," motivating deterministic quality gates, with Sonar suggested as a starting point for setting them up.
IntCyberDigest reports an active npm supply chain attack ('Shai-Hulud') that has compromised at least 868 packages carrying over 2 billion monthly installs with a credential-stealing worm . It began with the compromise of the GitHub account of the maintainer behind keyv (~127 million weekly npm downloads); a preinstall hook fires on npm install and drops a stealer that sweeps npm, GitHub, AWS, Kubernetes, and Vault secrets, then spreads to more maintainers . ThePrimeagen's response, 'did it ever really leave?', frames this as a recurring threat class . For agent-driven development, the attack triggers at the npm install step, making dependency installation the point to gate or audit.
swyx (@swyx) proposed a remote hackathon to clone the enterprise SaaS his team is about to pay >$40k/year for: he covers $1,000 in tokens per participant, contestants get a weekend to clone it, his team evals the result, the winner gets $10,000 cash plus a @latentspacepod writeup, and all code is open-sourced . The goal is to keep repeating this on increasingly ambitious SMB SaaS to map the boundary of what is still hard to kill in a weekend . A follow-up post says the event is live now, with Discord/briefing/rules via luma/check/calendar .
@swyx launched a $10,000 "kill my saas in a weekend" competition in which participants use any coding agent, any model, and up to $500 in token spend including subscriptions . The brief is live, the finish line was extended to Wednesday, and late entries can join a waitlist . Over 600 people applied, and 100 were admitted last night .
Anthropic will make Claude Code's auto mode the default for new sessions on Pro, Max, and Team plans starting August 14, 2026 (announcement: https://claude.com/blog/auto-mode-default-in-claude-code; docs: https://code.claude.com/docs/en/auto-mode-config).
Cat Wu (Anthropic) says nearly everyone at Anthropic uses auto mode and claims prompt-injection and data-exfiltration risks are "far lower than the average human reviewer" .
Eval with 1,053 paid testers: when a permission prompt was swapped for a dangerous command, only 13.6% of humans refused; auto mode would have blocked 89% of those actions, leaving ~11% unblocked .
Third-party eval (Trajectory Labs, commissioned by Anthropic) against the latest available Claude Code and Codex as of July 17, 2026, using 72 held-out indirect prompt-injection scenarios: 0 of 720 attacks succeeded against Claude Fable 5, Opus 5, and Sonnet 5 running auto mode .
Simon Willison's counter-signal: auto mode likely can't stop a malicious test dependency instructing the agent to run uvx fetch-model-files . then uv run pytest, where the first command exfiltrates data; he wants independent confirmation before trusting the claims .
He also notes confirmation fatigue makes human approval an unreliable safety mechanism and recommends running agents without access to data/tools that can cause harm if triggered .
@thsottiaux announced GPT-5.6 Sol is usable "pretty much anywhere, including in the CC harness," and reset usage limits for all paid ChatGPT Work and Codex users . @ThePrimeagen framed this as "the token fights are taking a weird turn" .
@agentnative_ published a video roundup of the past two weeks' frontier agent platform and model updates, shared by @rileybrown . Covered topics:
- Meta Muse Code, framed as a Claude Code competitor, including a setup walkthrough and "Yolo Mode"
- Codex updates: new sidebar, in-app browser, Chrome extension, Chat & Work toggle
- Cursor revamp, billed as becoming a "super app"
- New Chinese models
- "Anthropic's Silence" and what keeps the author on Anthropic
- "Buzz + Agent Teams" (teams of agents)
This is a secondhand roundup (a video chapter list), not a firsthand workflow report — the post itself contains no implementation details. Video: https://x.com/agentnative_/status/2086156650264633406
Anthropic's @trq212 announced auto-mode is becoming the default in Claude Code, in a claude.com post the author joked should be called "defeating the lethal trifecta" . Simon Willison (Django co-creator, PSF board) reacted with skepticism and published notes on his blog (simonwillison.net/2026/Aug/8/auto-mode/): he says he "REALLY" wants to believe auto-mode fixes prompt injection risks for coding agents, but he's "just not there yet" — a firsthand caution against trusting auto-mode as a prompt-injection safeguard .
- @swyx launched a live "$10,000 kill my saas in a weekend" competition: any coding agent, any model, up to $500 in token spend including subscriptions; winner gets $10k cash and a @latentspacepod writeup, all code open-sourced, and evals are run by his team (the prospective customer); the finish line was extended to Wednesday and the brief is live, with 50 people already started.
- Motivation: swyx's team is paying >$40k/year for an enterprise SaaS they have never used and cannot customize; he wants to repeat this competition with increasingly ambitious SMB SaaS to find the boundary of what is still hard to kill in a weekend. The original proposal covered $1000 in tokens per participant, while the live version caps spend at $500 including subscriptions.
@swyx calls Anthropic's ultracode "one of the most important coding mode innovations ever invented" . He urges developers to grasp the potential of dynamic workflows, citing a Kill My SaaS competitor who produced "a pretty good submission in 3 ultracode prompts" — a secondhand anecdote suggesting dynamic-workflow prompts can yield competitive output with very few iterations .
In a quoted post, @NickADobos argues the gaming category is dead because everyone will build their own games: "Games don't have ongoing maintenance cost like companies replacing SaaS vendors," AI brings creating a form and playing in it "closer and closer to zero," and "Just pay chatgpt $20 and make you own" . @ThePrimeagen counters from firsthand experience as a prominent developer: "Again, people have no idea how hard creating a good video game is... No, the gaming category is in fact not dead" — a contrarian caveat against the vibe-coding-can-replace-specialist-craft narrative.
@alexgetmancom (firsthand report) says Anthropic suspended his account shortly after he followed a published setup, linked to a post by @thsottiaux (Tibo), for running another model through Anthropic's coding harness via a proxy; he filed an appeal, asked whether the setup is actually prohibited, and made the full implementation public at https://github.com/alexgetmancom/claude-proxy. The referenced setup post: https://x.com/thsottiaux/status/2076119366647894371. Tibo replied that he doesn't work at Anthropic, said it "does seem odd that they would ban your account for using their harness with another model," and asked whether anyone else is in the same situation .
Riley Brown released a 61-minute tutorial, 'Learn 99% of ChatGPT Work in 61 minutes,' describing ChatGPT Work (referred to as 'GPT Work') as 'like Codex in the cloud,' available on phone, web, and desktop, and says he's been using it to run his business . The tutorial covers 14 knowledge work capabilities: Presentations, Plugins, Blocks, Websites & Apps & Hosting, Branching Chats, Desktop App (Cloud vs Local), Voice Mode, Remote Voice Mode, In-App Browser, Skills, Scheduled Automations, Spreadsheets, and a Multi-Agent Workspace, plus a Final Review (the timestamped list skips #5) .
ThePrimeagen reports he is "still at 10 prs merged a day from a cloud agent into projects" , then asks if that means he is "officially left behind" — a firsthand data point from a well-known developer on current agent-driven PR throughput, suggesting even ~10 agent-merged PRs/day may feel like a low bar as cloud agents scale up.
Marcos Hernanz ran ~1B tokens through OpenAI's GPT-5.6 Luna in a couple of hours . Romain Huet (OpenAI Head of Developer Experience) relayed that this came to roughly $30 — "Almost 1 billion tokens with GPT-5.6 Luna for $30. Incredible what you can do at this price!" . That implies ~$0.03 per million tokens, suggesting token-heavy agentic coding workflows (large context, many iterations) are becoming dramatically cheaper.
-
@reach_vb recommends regularly updating your
agents.mdand shares their own as a template: use the Visualize skill when explaining things; be concise, direct, and candid, challenge weak assumptions, and distinguish verified facts from uncertainty; ground research in authoritative, current sources and link important evidence; preserve the original goal and constraints, finish authorized work end to end, and verify the actual result before claiming completion; ask questions only when a decision is materially ambiguous, risky, or requires approval; use relevant skills and spawn subagents only for genuinely independent work, synthesizing their findings; keep changes focused — no unrelated edits, unnecessary abstractions, or low-signal tests; test observable behavior, review substantial changes, and validate user-facing work in the real interface; never take destructive, production, or external actions beyond what's authorized; report meaningful blockers, outcomes, and evidence without noisy progress . - Jediah Katz — whose bio says he's building the @cursor_ai agent — counters in reply: if you need to update agents.md regularly with guidance like "finish authorized work end to end," then AI developers have failed you .
Kent C. Dodds (creator/educator) says Cursor's cloud agents and bugbot are "the most critical part of my software factory," urging people to stop sleeping on Cursor — a firsthand production endorsement .
Debs Obrien's firsthand mobile workflow: while "nap trapped" in a car, she sent one prompt to Cursor's cloud agent via the mobile app to fix the login for her Playwright demo site. The agent spun up, streamed live updates to her phone, and she could review screenshots the agent built, then merge straight from the app or hand off to the GitHub app to review/merge. She called it "the best cloud agents experience" and "Cursor is winning this one hands down," noting she was not paid to say it .
