Summary

The same weekly GitHub round-up from The Next New Thing, but from 16 July 2026 — two months older than the other episode in this batch, and worth reading as the earlier snapshot. Andrew Warner and Adam cover three design repositories and the week's top ten, with a live interview with the creator of the week's number-one project. Commentary rather than testing: the presenters are open that most of these are things they have read about rather than run, and Adam repeatedly says he intends to try something rather than reporting that he did. One sponsor is read out. The episode's recurring theme, stated several times, is that managing several agents at once has become the hard part of the work.

• A problem stated plainly and worth recording as an observation about the tools rather than a product pitch: coding agents left to themselves produce designs that look like every other AI-generated design. Adam's phrasing is that you can tell at a glance which generator made a site, that fiddling with it does not really make it better, and that this is 'one of the most frustrating parts of building'.
• Adam's working practice in response is to install four or five different design skill packs and switch between them to get 'a bunch of different designers' perspectives' — treating the packs as sources of variation rather than as improvements.
• The most transferable advice in the episode, and it needs no tool: ask an agent for four divergent variants rather than one result. Adam's reasoning is that the agent is not especially creative and that the user often does not know what they want, so the value is in seeing options and combining parts of several — 'it's not about one of them being the best, it's about stimulating your creativity'. He specifically recommends using the word 'divergent' in the prompt.
• A candid failure report on diagram generation: Adam describes spending an evening unable to get a model to lay out a flowchart the way he wanted — it produced a twelve-step horizontal chart requiring heavy zooming, then a vertical one, and never the denser arrangement he was after, and he gave up. Warner notes a previously reported comparison finding one model better than another at this particular task and says he had assumed it was solved; Adam confirms he was using the supposedly better one.
• The week's top project is an AI job-search workflow, and its creator, Mads Lorentzen, appears live. He reports being laid off in mid-December while on parental leave, building the tool with Claude Code while the baby slept, and being hired using it — he confirms on air that he is now employed and used his own workflow to get there.
• The design principle he describes is the notable part: the tool evaluates fit against a posting and scores it honestly — he says he has declined jobs based on its assessment — and when it does draft an application it is 'grounded in your profile', explicitly not 'about making up the best or the perfect candidate'. That is a deliberate constraint against the obvious misuse.
• On measuring open-source success, Lorentzen makes a distinction worth keeping: forks matter more than stars 'because that's actually people using it'. He also volunteers unprompted that a lot of the stars are 'probably bots', which is a useful corrective to the star counts quoted throughout this series as evidence of significance.
• Adam pushes back on one of the week's popular repositories, arguing a desktop-control tool adds little because the models can already read and write files anywhere, use the terminal and find things themselves. His summary is the sharper point: 'eight months ago I would have been blown away, but now I think the built-in tools are doing this.' A reminder that tools built around a model's gaps expire as the gaps close.
• A screenshot-based trend is identified: giving models a visual view of the result rather than only the code. Adam says he does this naturally and funnels screenshots into his coding agents, and that a tool driving Word, Excel and PowerPoint works better for having 'eyes' — otherwise generated output places text over text and charts over charts.
• The cost argument for such wrappers is specific: a capable model can work out how to edit an Excel file unaided, but will spend a lot of tokens and time doing it. Adam adds that he has 'never seen it make a great PowerPoint'.
• 'Token maxxing' is named as a trend — deliberately spending the full token allowance across several subscriptions and providers at once. A related practice is described where a household member or colleague with a chat subscription they use only for chat lends the coding capacity they never touch. Reported as observed behaviour, no figures.
• A caution against automatic provider fallback that is more interesting than the tool: Adam warns against switching between different makers' models mid-project without realising, because 'I talk to them differently, they talk to me differently'. He endorses such tools for rotating between several accounts with the same provider, but says the user must always know which model is answering.
• The parallel-agent theme recurs across at least four separate entries in the episode — running several agents side by side with a worktree and its own running copy per agent, herding agents from one terminal, one endpoint fronting many providers, and cross-calling between two makers' tools. Adam notes the terminal gets features first and that 'probably half the features almost never make it to the extension anyway', which is why the tooling is converging on making the terminal manageable rather than moving to an editor.
• OpenAI is reported to have released a skill that lets Claude Code call out to Codex for tasks, with the definitions stating what it is good at and when to use it. Adam's reading is that every project should use more than one model and that this removes the burden of managing the parallelism by hand.
• A sandbox project is explained clearly and the underlying point is durable: if an application lets its own users drive an agent that can write and run code, that code cannot be allowed to run on the server, because it could reach the database, take other users' data or exhaust bandwidth. Isolation with constrained memory and no network or database access is the mechanism that makes user-facing agents safe.
• Adam remarks with evident pleasure that a set of Google C++ building blocks reached the top ten with nothing to do with AI, noting that for three years the list has been 'AI AI AI', and that people are still building the infrastructure everything else runs on.
• On video, the argument is that transcripts are an impoverished view: Warner notes most agent-based editors work from transcripts, which suits talking-head footage but misses what is on screen, and Adam adds that 'you don't see excitement in a transcript'. Both describe long recorded material they have never managed to repurpose because working through hours of video by hand is impractical.

Why it matters