We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Quality comes from checks the agent can run
Addy Osmani made the clearest argument this period. He says agents write better code "at scale" because they make checking cheap. Tests, benchmarks, specs and real end-to-end runs used to be what made good code expensive, and now you can ask an agent for them. So the job shifts from "review every line" to "build loops the agent can run itself" . He also says to make taste compound: when the agent gets something wrong, write the fix into shared config, a hook or a test so it doesn't happen twice. You still own what ships, but you review evidence as well as diffs .
He was responding to a complaint that asking Claude to review code is a slot machine that finds new bugs every time. His fix for that has three steps :
- Define "done" before review starts. Without that, a reviewer (human or agent) can always find something to improve.
- Give Claude checks it can run, such as bash commands, tests or clicking through the app in a browser, and require them to pass before it says it's finished.
- When nothing blocks, merge and ship.
Salvatore Sanfilippo (antirez) shows the same idea on hardware, in a video where Opus 5.5 reworks his RP2040 ZX Spectrum emulator and adds a Commodore 64 emulator. The video is in Italian. Several steps are worth copying:
- Make a separate scratch copy first. Before porting his C64 emulator into the device project, he had the agent make a gitignored, uncommitted copy inside the repo. It converted that copy to instruction-at-a-time execution on its own and reported back before any integration. He warned it would get tangled if it converted and ported at the same time .
- Use the old implementation as the oracle. The agent ran the old cycle-accurate emulator and the new one side by side, requiring matching images and registers. After 10,000 frames the output was still bit-identical .
- Measure, then redesign. The faithful port ran at about 10% of real speed. The bottleneck was the VIC video chip . He told the agent to capture register writes and render frames from them instead of emulating the chip cycle by cycle . The result runs Spectrum and C64 in one device at 400 MHz, including timing-heavy C64 demos at 50 FPS with sound .
- Give the agent instrumentation and inputs. It simulated key presses itself . He says its first port worked on the device first try , and its instrumentation checked for delays, dropped frames and audio problems . His C64 emulator also exposes a direct API for stepping, debugging and grabbing frames, which he prefers to an MCP layer he sees as adding friction .
Separately, MaciekCodes reports that his own coding agent, "a 'while loop' and a few tools," improved itself quickly .
GPT-6.1 Sol: bigger default context in Codex, and plan math
Tibo (OpenAI) says GPT-6.1 Sol's default context in Codex & Work went from 272k to 500k, about 80% more. The aim is fewer compactions and better handling of long, nuance-heavy tasks. You can set it higher or lower, up to Sol's 1M maximum . Subscription usage of Sol carries no long-context fees .
Theo thinks the default speed was raised too much and is now barely slower than "fast" . He estimates he could use Sol Ultrafast as his main model on the $500 plan and only just hit his limits . That figure is his total usage (583/week), and it leaves out his long-running ts-rust work . He also built a dashboard that tracks Claude usage across his accounts to roughly predict how long his limits will last .
Theo: judge models by cost per completed task
In a new video, Theo argues that per-token price is the wrong measure for agentic work. Token-hungry models can wipe out a low sticker price. For agentic dev work he says caching, reliability and token efficiency matter much more . He says bills are rising because agents now run longer and finish bigger tasks . He favors GLM 5.3 Flash for real-world agentic work, citing its vision, tool calls and long-running work .
The sponsored segment of the video (Railway) still describes a usable pattern. On his Lakebed project, agents set up production, staging and per-PR environments for him. Claude or Codex connects through Railway's MCP, spins up an environment, verifies a change, ships by merging to main, then checks on the environment afterward .
For local compute limits, he moved ts-rust's build and test commands onto Blacksmith runners. Each thread still takes a while, but he is far less bottlenecked when running many in parallel .
T3 Code nightly
Theo added a proper /goal flow with a visual callout to the T3 Code nightly . A user showed off the new inline rendering in chat by asking the agent whether a monitor mount would clamp onto their desk . Theo says he hadn't expected the feature to be so useful . For hazy features he still uses Matt's "grilling" skill, noting that modern models are good at asking questions .
Smaller notes
-
Riley Brown shared a
grokbot-webhookskill. Claude Code hands X/Twitter research to GrokBot, waits for the answer, then continues with the task . He says it works with any agent, including Codex . - Kent C. Dodds shared the Kody v2026.10.11 release (release notes). Cron jobs, webhooks, subscriptions and inbound email now run with the permissions of whoever enabled them, checked live on every run, with no fallback to the org Owner . His caption: "This is what 122 PRs in one day looks like" .
- Armin Ronacher says no single agent runtime wins. VMs, sandboxes, durable objects or a plain JavaScript interpreter each fit different situations .
- For his RP2040 ZX Spectrum emulator, Salvatore Sanfilippo asked Opus 5.5 to improve an existing project using changes he had in mind; the work included writing directly to Spectrum video memory and using an instruction-accurate Z80 emulator instead of the previous cycle-accurate approach. Opus simulated button presses, and Sanfilippo reports that the first device-tested port worked on the first try; a later scanline fix stopped BMX sprites from flickering in about five minutes. He also used on-device checks for delays, dropped frames, audio, and video corruption: at 266–280 MHz, the emulator maintained 50 FPS without visible corruption, and he described the audio as effectively glitch-free.
- For the C64 port, he instructed the agent to make a gitignored, uncommitted copy of the existing emulator and convert it separately before integrating it, moving from cycle-by-cycle execution toward one instruction per step. He compared the old and converted versions over 10,000 frames; the output remained bit-identical. When the port initially ran at only 10% of real speed, they identified VIC video rendering as the bottleneck and replaced internal cycle-by-cycle chip emulation with register-write capture and rendering. The resulting device ran both Spectrum and C64 emulation at 400 MHz, including timing-sensitive C64 demos at 50 FPS with sound.
- Sanfilippo describes his C64 emulator as offering a direct LLM-facing API for stepping, debugging, and capturing frames, avoiding an MCP layer between the model and the software.
- Karpathy says frontier coding agents can complete large, complex projects from a prompt that previously took weeks or months, including building, copying, or translating apps and decompiling binaries. He names OpenAI Codex and Claude Code, and describes Codex working for an hour to restructure an entire codebase.
- For task selection, he argues that impact grows with a problem’s ambition, size, and time horizon, while short-answer questions barely test these systems; he says capability is strongest in digital, verifiable, economically valuable work, giving unit-test pass/fail as an example of a verifiable result.
-
On his Lakebed project, Theo uses Railway’s skills, CLI, and MCP to let agents provision and manage the app. With Claude or Codex connected through MCP, an agent can investigate a requested change in a new environment, verify it, ship by merging to
main, and check the environment afterward; his setup includes production, staging, and PR-specific test environments. - For agentic development, Theo recommends looking beyond token prices: token-hungry models can erase apparent per-token savings, so caching, reliability, token efficiency, and cost per completed task matter. He says higher bills also reflect agents running longer and completing bigger tasks.
- Theo favors GLM 5.3 Flash for real-world agentic work, citing its vision, tool-calling, and long-running-work capabilities; he argues that strong tool use can let smaller models outperform older, more knowledgeable models.
For tedious, repetitive game decompilation, Casey argues AI could help by learning mappings from source code to compiled output; reverse engineers already use machine-assisted tools such as IDA Pro and Radare, so AI would extend an existing tool-assisted workflow. AI-generated decompilations can be poor representations of the original source: the panel describes the Banjo-Tooie output as atrocious and of limited value for understanding or preservation. Its practical role may be a time-saving starting point that a human cleans up, rather than a finished artifact.
Kent C. Dodds linked Kody v2026.10.11 under the caption “This is what 122 PRs in one day looks like”; the post does not specify that Kody produced those PRs. The release makes cron jobs, webhooks, subscriptions, and inbound email run with the permissions of the person who enabled each trigger, checked live on every run, with no fallback to the organization owner.
Osmani argues that agents can improve code quality at scale when verification is cheap: build loops the agent can run itself, using checks such as tests, benchmarks, specs, and real end-to-end runs, rather than reviewing every line . When an agent makes a mistake, encode the correction in shared config, a hook, or a test; the human still owns what ships, but can review evidence as well as diffs .
Define “done” before asking Claude or a human to review code; without that boundary, a reviewer may keep finding things that could be improved. Give Claude runnable checks—such as bash commands, tests, or browser walkthroughs—and require them to pass before it declares the work finished; merge and ship when nothing blocks it.
A grokbot-webhook skill lets Claude Code delegate X research to GrokBot, which has full access to X posts; Claude waits for GrokBot’s answer, then continues the requested task. The author says the workflow also works with other agents, naming Codex, Muse, and Instinct.
MaciekCodes reports that his own coding agent—a simple while loop plus a few tools—improved itself quickly, and that models are already effective with this basic toolset . Geoffrey Huntley’s takeaway is to use that knowledge to build “recursive products” .
Claude Code can delegate X/Twitter research to GrokBot through the grokbot-webhook skill: GrokBot has access to X posts, returns an answer, and Claude then continues the requested task. An AgentNative setup guide explains how to connect Claude Code to GrokBot with a webhook: GrokBot–Claude Code webhook guide.
In response to a call for software that works without three hours of prompting, Riley Brown argues that products promptable into existence that quickly will become lab-native “artifact types.” He cites Anthropic Motion for collaborative motion-graphic editing, GPT sites with built-in storage, authentication, databases, and hosting, and Claude Design’s editable Figma-style boards; he predicts OpenAI GPT Space could expand from pages and sites to any company file type.
Theo estimates he could use GPT-6.1 Sol Ultrafast as his main model on the $500 plan while just barely staying within its limits . He excluded ts-rust from the day-to-day usage behind the estimate, describing it as a long-running YOLO mission .
Armin Ronacher floated putting “effect” into “codemode” , in response to praise for Pi Durable’s “durable execution” .
Riley Brown says Codex computer use is “SO good” and that he “can’t live without it now,” a strong positive firsthand endorsement without further workflow details.
T3 Code nightly now includes a proper /goal flow and visual callout, making goal-setting a more explicit part of the coding workflow; the post gives no further usage details.
Theo moved ts-rust build and test commands from his machines to Blacksmith runners; he says this reduced his bottleneck when working on many threads in parallel, although each thread still takes a while.
For ts-rust, Theo moved all build and test commands from his machines to Blacksmith runners; although each thread still takes time, he says this made him less compute-bottlenecked when working on many things in parallel .
Tibo said the poll results surprised him: he expected stronger support among Pro 100/200/500 users for the “2X” direction, even if it meant bringing back five-hour limits to smooth demand. This reflects his expectation about user preferences, not an announced product change.
A post claimed someone was building an installable Spider-Man 2 mod that brings in Skate 3’s skating mechanics, calling it AI-powered modding . Theo disputed the claim, saying foot placement alone made him think everything in the post was false .
Theo recommends switching to Linux to address the heat reported when running 20–30 subagents on a MacBook with an M5 Pro and 48 GB of RAM; the recommendation is brief, with no setup details or comparison provided.
