We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
Benchmark the harness, not only the model. A new Latent Space analysis reports the same model scoring 52.4–76.2 on the same 106 tasks across different harnesses—a 23.8-point spread—and GPT-5.6 Sol’s ARC-AGI-3 score rising from 13.3% to 38.3% through retained reasoning and compaction alone.
The practical consequence: treat context, memory/compaction, tools, permissions, and guardrails as model-adjacent engineering surfaces, not wrapper polish. The analysis argues that as models absorb more of the scaffold, the remaining product is an “attention-interface” governing when an agent interrupts, what it can decide alone, and what needs approval.
⚡ TRY THIS
Run an async, bounded overnight loop. DHH gave Omabot a HEY account and beta CLI, kicked off work, and woke up to an email describing what it had completed; he expected to ship the result to edge that day. Replicate the pattern on a reversible task: provide a scoped account/CLI, define the deliverable and completion channel, let the agent work asynchronously, then review the output before shipping.
Use one painful surface as a three-prompt smoke test. Theo first said he could build something better than the Plex app in under five prompts, then followed up with “3 prompts,” streaming from his NAS over Tailscale. Pick one irritating personal workflow, preserve the existing data path, ask for a working end-to-end slice, and cap the experiment before architecture work takes over.
Parallelize only after the spec is legible. Matt Pocock’s in-progress
/implement-specskill takes a spec and tickets, researches the codebase in a subagent, implements tickets in maximally concurrent subagents, reviews against the spec, and cleans up worktrees. Kent C. Dodds supplies the needed counterweight: being overly prescriptive can undo optimizations made by the model and harness. Use the phases as contracts and outcomes, not as a script for every subagent move.Make verification behavioral—and do not accept the first “impossible.” Simon Willison frames productive agent use as confidently instructing the change and then confidently verifying it, rather than assuming line-by-line eyeballing is the best validation method. Linus Torvalds describes the concrete debugging version: after an AI repeatedly declared a kernel problem unsolvable, he pushed it to keep adding debug code and analyzing faithfully, then let it write the commit message. Ask for instrumentation and evidence before treating a refusal as a conclusion.
📡 WHAT SHIPPED
llm 0.33— Simon Willison’s CLI upgraded to the OpenAI Python library 3.x andhttpx2; embeddings now accept per-call--key/key=values. Repeated-t/--templateflags now compose saved model configuration and prompts, including this directly usable pattern:llm -m gpt-5.6-luna -o reasoning_effort high --save lhigh llm "Generate an SVG of a pelican riding a bicycle" --save pelican llm -t lhigh -t pelicanReasoning-capable Responses API models also expose
reasoning_summary=auto|concise|detailed.Codex CLI startup was rebuilt for latency. Charlie Marsh says the latest release redesigned the lifecycle to make
codexstartup instant, roughly 25× faster and immediately responsive.Kody Koala is positioning itself as a cross-agent runtime. Kent C. Dodds says it augments, rather than competes with, tools including Cursor, Codex, Claude, OpenCode, OpenClaw, and Devin, with work portable across them.
Vision is becoming a practitioner selection bar. Theo’s explicit view—not benchmark data—is that a code model unable to inspect a screenshot of an issue is “wholly disinteresting”; he calls vision a base-level requirement.
🎬 GO DEEPER
▶️ DHH’s Linux/Omarchy video — 10:20. DHH linked this exact timestamp while arguing that “Linux is fucking happening”; use it as the visual companion to the async Omabot workflow above.
The Evolution of the Agent Harness. Read the Harness-Bench comparison alongside the proposed attention-interface: the useful question is not merely which model wins, but which context, compaction, permission, and approval design lets the same model do reliable work.
Armin Ronacher — Fast, Hard Code. Ronacher argues that agents reduce the friction of unfamiliar languages, points to
pi-autoresearchfor agent-directed performance work, and notes LLM-assisted projects in “hard” languages such as Zig. The useful caveat is his own: domain knowledge still helps.Study the
implement-specskill. It is a compact reference for turning a spec into codebase research, concurrent ticket execution, a spec review, and worktree cleanup—while the Kent warning above argues for leaving the harness room to optimize.
Editorial take: The frontier is shifting from giving agents more autonomy to engineering the boundary around them—preserve context, delegate asynchronously, parallelize bounded work, and spend scarce human attention only where the agent cannot decide alone.
The essay provides no concrete prompts, workflows, project-structure or verification techniques; it is high-level commentary on LLMs lowering language-learning friction and enabling agent-assisted performance work. Relevant signals:
- Language-learning no longer matters as much for agents; programmers can pick languages they are unfamiliar with and have the agent work in them
- Language choice is increasingly based on marketing/perception, and LLMs are described as "exceptional at optimizing code without regressing behavior"
-
Performance optimization can be delegated to an agent via tools like
autoresearch, though "knowledge greatly helps" -
Example LLM-assisted projects: Cloudflare Artifacts (pure-Zig Git-protocol engine, ~100 KB WebAssembly module) and Vercel
fx(a Zig coding agent advertised as small and fast); "all these projects are largely LLM-assisted" - Agents are enabling work on "harder" technologies previously off-limits (DWARF files, eBPF, custom network drivers, custom crypto, old hardware)
- Caution: the author expects "more slop" along with more developers wanting fast/small software; no operational cautions are given
- Simon Willison (22 Aug 2026) argues the core skill for productive coding-agent use is confidently instructing an agent how to make changes and then confidently verifying those changes were applied correctly — and that reviewing every line of agent-written code is not the most effective validation method, since eyeballing code line-by-line was never the best way to validate a software change.
- His blog's recent-articles list flags Qwen 3.8 27B (16 Aug 2026) as "excellent, but it defaults to wildly overthinking things" — a useful signal for coding-agent model choice and routing when excessive reasoning overhead hurts.
- Sponsored note from Teleport: AI agents "don't sleep and will try anything to achieve their goal," so deploy them safely in an isolated ephemeral trusted runtime.
- Harness impact is measurable and large: the same model scored 52.4–76.2 (23.8-point spread) across harnesses on Harness-Bench's 106 tasks , and OpenAI tripled GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% with only retained reasoning and compaction .
- Claude Code's design choices worth replicating: terminal-based, gives the model bash and file read/write access, replaces per-change human approval with permission rules, and was built with the next model's capabilities in mind — reaching ~$1B ARR in six months .
- Anthropic deleted 80% of Claude Code's system prompt ; the loop is train → absorb → shed → repeat , so the measure of harness evolution is how much you can delete while retaining capability .
- Autonomy without capability fails: 95% per-step reliability over a 20-step task yields only ~36% average success, making early AutoGPT/BabyAGI-style full autonomy premature .
- Next pattern to build for: an "attention-interface" — a per-user policy surface (analogous to AGENTS.md for codebases) governing when the agent may interrupt, which decisions need approval, and what it can decide alone; predicted to ship at every agentic AI company within a year, with corrections becoming training data .
Linus Torvalds described a Linux kernel 'debug session from hell' (drm/xe driver) that was 'enormously helped by an AI doing much of the grunt-work.' The unnamed AI repeatedly claimed the problem was impossible/unsolvable and suggested writing a report, but when Torvalds pushed it, it kept adding debug code and analyzing faithfully; he let the AI write the commit message. He joked the AI seems 'trained by people who may not be quite as stubborn as I am.' Firsthand account of production kernel debugging: the practical pattern is to keep an agent working on instrumentation/analysis even when it signals defeat, and to use it for commit message drafting. Commit: 818bebeb
X user @GeoffreyHuntley asserts that in 2025 LLMs became good enough to displace JetBrains/IDEs, and that in 2026, demonstrable experience automating your own job function and the job functions of your colleagues became the new norm for getting a job . He offers no tools, benchmarks, or workflow specifics, so this reads as a contrarian industry claim about IDE displacement and hiring norms rather than a firsthand usage report; the post includes a video whose content is not described in the text .
Ben Tossell (@bentossell), a builder and dev-tools/infra investor, says he never updates Chrome but always updates his agent apps — a first-hand signal that coding-agent tools have become a more essential part of his daily workflow than the browser, though no specific tools, workflows, or techniques are named .
Simon Willison released llm 0.33, his CLI for accessing LLMs, upgrading to OpenAI Python library 3.x and switching the HTTP client dependency from httpx to httpx2.
llm prompt -t/--templatecan now be repeated to combine templates in order — enabling a practical pattern where a model config with default options is saved separately from a prompt, then invoked together:llm -m gpt-5.6-luna -o reasoning_effort high --save lhigh,llm "Generate an SVG of a pelican riding a bicycle" --save pelican, thenllm -t lhigh -t pelican.llm embedandllm embed-multinow accept--key, and the PythonEmbeddingModel.embed()/embed_multi()andCollection.embed()/embed_multi()methods acceptkey=, matching the per-call key pattern already used by LLM models; existing plugins readingself.keykeep working via a compatibility fallback .-
Reasoning-capable Responses API models now support a
reasoning_summaryoption withauto,concise, anddetailedvalues, usable viallm openai endpoint --responses.
DHH (@dhh) credits what Asahi has achieved but suggests 'SOMEONE' should 'spin up a new effort, lean in with the best agents, and actually finish the job on M3, M4, M5' — an opinion, not a firsthand workflow, that top coding agents are capable of hard systems-level porting work. It names no specific agents, tools, or steps.
- Matt Pocock is testing an in-progress
/implement-specskill — a multi-agent implementer that takes a spec and tickets, does codebase research in a subagent, implements all tickets in subagents with maximum concurrency, reviews the final code against the spec, then cleans up all worktrees. Goal: "smash out huge chunks of work autonomously with minimal supervision." Skill file:https://github.com/mattpocock/skills/blob/main/skills/in-progress/implement-spec/SKILL.md - Kent C. Dodds's contrarian take: being overly prescriptive about workflows like this is a good way to undo optimizations made by the model and the harness — i.e., too much rigid orchestration may degrade results versus letting the model/harness optimize.
Kent C. Dodds (@kentcdodds) highlights Kody Koala as a runtime that augments rather than competes with coding agents @bot, Cursor, ChatGPT/Codex, Claude, OpenCode, OpenClaw, and Devin, and makes the work developers do portable across all of them .
After saying the Plex app is so bad he was tempted to build a clone and was "pretty sure I could make something better in under 5 prompts" , @theo followed up that he did it in 3 prompts, streaming straight from his NAS over Tailscale . He shared a video of the result .
T3 Code, an agent-running app, has proven highly addictive for its users: @linuz90 deleted it from his phone because he "couldn’t resist spinning agents all day long, even when out with friends" and cited mental health as priority . @theo said he has "had to delete it multiple times too lol, genuinely addicting" and that "Showers are no longer peaceful, I just stand in the water managing agents" . Both are firsthand accounts.
Codex CLI's latest release redesigned its startup lifecycle, making codex launch ~25x faster and immediately responsive — reported directly by the developer behind the change, Charlie Marsh . Flask creator Armin Ronacher (@mitsuhiko) highlighted it as an example of 'some of the best engineers with agency' shipping good work .
Theo (@theo, t3.gg) posted a video titled "I ranked every AI model": https://video.twimg.com/amplify_video/2091277559287996416/vid/avc1/3840x2160/l72NIfLWVhoMpn2O.mp4?tag=29
Developer Theo (@theo) posted a "rough tier list" ranking where he'd put every major AI model right now, alongside a video edition titled "I ranked every AI model." When asked where Ox Alpha ranks, he replied that it depends on size and price.
Armin Ronacher (@mitsuhiko, Flask creator) shared a link to his weekend essay on how LLMs change the way developers start new projects . The post points to https://lucumr.pocoo.org/2026/8/22/fast-hard-code/.
Omarchy (agentic coding) runs on an 11-year-old PC: @morganlinton's firsthand report says the first tweet from that machine came from Omarchy, it "feels faster than my two year old MacBook Pro", installed in 3.5 minutes, and he was "actually doing agentic coding, my normal workflow, immediately" . DHH reinforces that no new hardware is needed, warning it might cause buyer's remorse on a new MacBook Pro .
DHH (Ruby on Rails creator, 37signals CTO) reports a firsthand production workflow with Omabot, an AI coding agent: he gave it work overnight, it emailed him with what it had completed by morning, and the results were shipped to edge the same day . He set up Omabot with Ori as an agent option and connected it to OpenRouter (via its CEO Alex Atallah, after a suggestion from Patrick Collison) . In a follow-up he clarifies the target is HEY: he gave Omabot a HEY account and the beta CLI, and "It loves it" . The workflow is a hands-off overnight agent loop — assign a task, agent works autonomously, emails a morning report for human review, then ship — a concrete human-in-the-loop orchestration pattern a developer could replicate.
Flask creator Armin Ronacher (@mitsuhiko) reports that, thanks to his coding agent, he now regularly writes more Perl than Ruby and PHP — a firsthand, contrarian observation about how agents can shift language choices .
Theo (@theo, CEO of t3dotchat) argues that code models without vision capability — i.e., unable to look at a screenshot of an issue — are "wholly disinteresting," calling vision a "base level requirement" for coding models . This signals a practitioner expectation that coding-agent models should support multimodal issue input. Firsthand opinion, not benchmark data.
