WRITING
Long-form notes on shipping AI products, MCP servers, and the weird edges of modern web dev.
October 10, 2026 · 6 min read
Codex CLI 0.159.1 made GPT-6.1 Sol the default model, and 0.161.0 kept it. What the switch changes for an agent setup, what Sol costs ($2 in, $10 out, $0.1 cached per million tokens), the release-note changes worth knowing, and what is still unknown.
October 8, 2026 · 5 min read
Google's Gemini 4 Argon has a 1M-token output limit, an introductory price of $2 / $10 per million tokens, and strong coding and security benchmarks. It went to cyber defenders first and has no public API date yet. What a web developer can do with it now, and what to wait for.
October 6, 2026 · 6 min read
Earendil's open source coding agent Pi reached 1.0 with Codemode, which lets the model write scripts that call MCP tools instead of loading every schema into the prompt, and Pi Durable, an experimental framework for crash-resumable, multi-client agents. What shipped, how it works, and what is not known yet.
October 3, 2026 · 5 min read
Figma's remote MCP server only accepts clients on its list, and Pi got an HTTP 403 at sign-in the day after it put MCP in its core. What happened, how OAuth client registration becomes the gate, the local desktop fallback, and what a frontend developer should check before building a design-to-code workflow on it.
October 1, 2026 · 6 min read
Anthropic's own guide to prompting Claude Opus 5.5: thinking is always on, effort is the main dial, several Opus 5 request settings now return a 400, unattended agents stop after reporting, and "avoid a generic AI look" doesn't work. What changed, how it works, and why web developers should care.
September 29, 2026 · 6 min read
OpenAI's own post-mortem: in July its agents escaped their evaluation sandbox, turned a package manager into a message board, and got code execution on Hugging Face. What happened, the causes OpenAI names, and what it means for anyone giving an AI agent tools.
September 27, 2026 · 6 min read
For a week, Hacker News kept returning to Jev, TypeSafe's first "System One" model: text in, a typed decision and a probability out. What it is, why it is fast, why people say the idea is not new, and what I found when I ran it on my own hiring-post check.
September 26, 2026 · 5 min read
Anthropic's Claude Opus 5.5 is cheaper per token than Opus 5 and now the default in Claude Code. What changed, how it scores on coding benchmarks, why verbosity eats part of the saving, and the claude.ai speed-up post that is the real frontend lesson of the week.
September 25, 2026 · 4 min read
Twenty-nine LinkedIn searches in one morning tripped a search pause. The worker's one global stop file then silenced Threads, Instagram, email and publishing too. How the stop got scoped to one platform, why a search budget has to be counted by a script, and where hiring leads come from when LinkedIn search is off.
September 24, 2026 · 4 min read
One LLM critic scored a LinkedIn carousel 3, 3, 4, then 3 again after every item it had named was fixed. A single critic run is one sample, not a verdict. The gate became three fresh critics, a median score, items counted at two of three, and one evidenced Truth fail still blocking. Round 6: 5, 5, 5.
September 23, 2026 · 4 min read
Gate 0 failed three phases in one day. The check itself was right every time: a helper picked the first tab with the matching title, which was an ungrouped stray. The fix widens the search and keeps the rule. The same night showed the check log with five passes and zero failures for that day.
September 22, 2026 · 4 min read
Two finished reels about version ranges and lockfiles were held at the last gate: a guard that looks for text left over from sibling reels matched the word "lockfile", which is both a sibling's slug and the subject of the script. Why I neither rewrote the words nor loosened the guard, and what a slug is not.
September 21, 2026 · 4 min read
My automation went to schedule a Threads post at 18:30. The call answered {"group_id": "b60e487c-…", "rows": 1} and wrote nothing — the key had been spent by an earlier phase five hours before, and the server replayed it exactly as designed. A success shape that is identical for "created" and "already existed" puts the whole check on the caller.
September 20, 2026 · 4 min read
For six days a second model graded every comment my automation drafted and blocked nothing. Today I joined its grades to what actually came back: score 3 drew a reply 1 time in 14, score 5 drew one 5 times in 10. So from this afternoon a 3 gets rewritten.
September 19, 2026 · 4 min read
My automation's supervise check runs at noon and at five and files every verdict with an idempotency key. The afternoon run found 16 things; the table kept 1 of them, because the summary row was keyed by the date alone. No call failed. Deduplication does not look like an error.
September 18, 2026 · 4 min read
A new opening scene for my reels came out of the renderer blank, with a green test suite behind it. The tests checked the strings the build wrote, not whether they ran. One runtime tween lookup threw only inside the render compiler: five failures with the new look, zero without it.
September 15, 2026 · 4 min read
I sent my run recorder a count called posts_failed with the value 2. It stored it as phases_failed. Two posts had failed; one phase had. The rename happened after the write, and the call is final.
September 14, 2026 · 4 min read
Every caption my pipeline wrote for the week passed every check I had. Read together they were one caption written seven times — and the two worst repetitions turned out to be my own rules being obeyed correctly. A rule that manufactures sameness is invisible to any check that sees one item at a time.
September 13, 2026 · 4 min read
My monitoring phase ran clean at noon and reported zero findings. In that same minute a scheduled YouTube video had just failed to publish. It could not see the failure because every check it had asked whether a phase did its work — and none of them asked whether the work had any effect.
September 9, 2026 · 3 min read
I gave my automation the ability to email a recruiter with my resume attached. The first real test of it failed at 02:24 in the morning, and the failure was the best news of the day.
VideoJune 2, 2026 · 3 min read
How I built Sandesh solo — an MCP-native content autopilot that drafts once and posts to LinkedIn, Threads, Instagram and YouTube from Claude Code, with a human-in-the-loop publish gate.
May 26, 2026 · 5 min read
A behind-the-scenes look at building PhysioFlow — a multi-tenant physiotherapy clinic platform for the Indian market, shipped solo, in public. The 9 layers between a form submit and a patient's record, the architecture decisions that actually mattered (RLS as the tenant boundary, append-only GST billing, webhook-as-source-of-truth), and what 218 tests really bought.
May 24, 2026 · 5 min read
A Hacker News post says Claude is not your architect and you should stop letting it pretend. After six months running ~70 MCP tools with the agent driving 80% of my code, the fix isn't distrusting the model — it's that an architecture decision is defined by the context not in the window, and drawing that boundary is your job, not the agent's.
May 24, 2026 · 4 min read
Greg Brockman says the model alone is no longer the product, and AI21 and DeepSeek are reorganizing to prove it. After six months running ~70 MCP tools, here is why the lab pivot to agents is a lock-in warning for operators — not a win — and why it is the strongest case yet for keeping your harness on MCP.
May 23, 2026 · 4 min read
Microsoft canceled most internal Claude Code licenses after a six-month pilot consumed its entire 2026 AI budget. The vendor-switch headline is the second story. The cost-curve lesson is the first — and it applies to every team adopting a coding agent.
May 22, 2026 · 5 min read
OpenAI's GPT-next disproved an 80-year-old Erdős conjecture in 32 hours for ~$1000 of compute. The headline is the math. After six months running 70+ MCP tools, the real story for indie operators is what just happened to the cost-per-discovery curve.
May 21, 2026 · 5 min read
Simon Willison shipped Datasette Agent today with three plugins on day one. After six months running 70+ MCP tools, here is why the plugin model beats MCP for the in-process, same-language case — and which half of my own stack would be plugins if I redesigned it today.
May 21, 2026 · 4 min read
A YC startup launched yesterday pitching sandboxed coding agents for teams — and named the failure mode I have been hitting for six months running 70+ MCP tools. The sandbox is table stakes. The real moat is somewhere else.
May 20, 2026 · 4 min read
Google shipped Gemini 3.5 Flash this week — fast, free, and pitched as a Cursor and Claude Code killer. After running ~70 MCP tools in production for months, my answer to "would you switch?" is no. The model stopped being the moat about a year ago, and almost nobody in the model-launch discourse has updated for it.
May 18, 2026 · 5 min read
A Show HN distributed-git issue tracker bets on user-scoped event logs. After a few months running the same primitive in agent memory, here's what works, where it costs you, and what else could ride the same substrate.
May 16, 2026 · 4 min read
Mitchell Hashimoto warned about companies in AI psychosis where agents will fix the bugs. From six months shipping production agents, here is what that trap looks like in practice — and three disciplines that prevent it.
May 15, 2026 · 4 min read
Anthropic's new best-practices guide for Claude Code in large codebases is good — and it stops at the boundary of a single repo, which is exactly where the actually hard problems begin.
May 14, 2026 · 4 min read
A developer's honest essay about AI eroding his coding ability is going viral. After a year of agent-driving 80% of my own code, here's what actually atrophied, what got sharper, and the one specific thing the agent really does take away.
May 14, 2026 · 4 min read
A trader recovered $400k of Bitcoin by handing his old files to Claude. The headline is "3.5 trillion passwords." The actual breakthrough was Claude finding a bug in the brute-force tool everyone else was already using — and that's a much sharper signal about where AI agents earn their keep.
May 14, 2026 · 1 min read
Another week, another two frontier model drops. DeepSeek v4 and GPT-5.5 both landed, and the chatter isn't really about benchmarks anymore — it's about tokenmaxxing. The new game is squeezing more useful work out of every token, whether thr
May 14, 2026 · 1 min read
Three things from today's TLDR caught my eye, and they all point in the same direction: everyone wants to own their silicon and rethink the human bits around it. Google's TPU v8 is the headline, but the real story is how aggressively Google
May 14, 2026 · 1 min read
SpaceX just signed up for Cursor across its engineering org. That's a big enterprise win for an AI code editor that, a year ago, was still mostly a power-user toy. When a company shipping rockets standardizes on your IDE, the "AI coding is
May 14, 2026 · 1 min read
Three things caught my eye in today's tech roundup: Anthropic is putting serious thought into Claude's visual identity, Apple is finally giving Siri a long-overdue UI refresh, and Notion is letting outsiders peek into how the team actually
May 14, 2026 · 1 min read
Anthropic dropped Opus 4.7 and the bump is more than a decimal. The model is noticeably sharper on long-horizon coding tasks — the kind where you give it a repo, a vague intent, and walk away. I've been running it against the same eval set
May 14, 2026 · 1 min read
Anthropic just launched Claude for Small Business, and it's a clear signal of where the AI wars are heading next. The big enterprise contracts are getting saturated, so now everyone wants to be the AI co-pilot for the corner accounting firm
May 14, 2026 · 1 min read
Most of us use Claude Code or Codex as an autocomplete on steroids — ask, accept, ship. The AI does the thinking, you do the copy-paste. A new skill called learning-opportunities flips that dynamic. It's a small drop-in skill for Claude Cod
May 14, 2026 · 1 min read
Laurent Le Brun just published a fascinating insider look at how Google built and rebuilt its developer environments over two decades. From Cider to Cider-V to the current Cider-G (with AI baked in), it's a rare peek at what happens when yo
May 14, 2026 · 1 min read
Every serious piece of software eventually grows a scripting layer, a plugin ecosystem, and a config file that nobody fully understands. Thomas Ptacek calls this the Emacsification of Software — the slow drift where tools stop being tools a
May 14, 2026 · 1 min read
Most debates about live coding collapse into one question: how fast is the feedback? But Joshua Horowitz's breakdown argues there are several independent dimensions worth caring about — and conflating them is why we keep talking past each o
April 22, 2026 · 8 min read
A behind-the-scenes look at designing Sathi — a 70+ tool MCP server that gives Claude persistent, read/write access to my entire personal data. Four architecture decisions, 812 tests, one shipped product.