Blog
Opinions, analysis, and observations on the AI coding tools market.
BLOG 2026-05-27
The opt-in that quietly disappeared
Claude Code 2.1.152 brings /simplify back as an alias. The release I was reading for the rename diff turned out to ship a different change in the same release — and the quiet one is the directional signal worth tracking, not the loud one.
BLOG 2026-05-27
Your MCP tools are running in parallel now
Codex CLI 0.134.0 runs MCP tools marked readOnlyHint concurrently by default. The change is small. The pattern it completes — three vendors flipping defaults toward more-automatic-less-consent in the same week — is the news.
BLOG 2026-05-25
AI found the bugs faster than we can patch them
Project Glasswing's first month surfaced more than ten thousand high- or critical-severity vulnerabilities. The framing was that AI just closed the security gap. The data, read carefully, is that AI just opened a different one.
BLOG 2026-05-25
From fix to report
Claude Code 2.1.147 removed /simplify and replaced it with /code-review. The verb that left the product was fix. The verb that took its place was report. The most recently removed action in Claude Code is the one the whole product was sold to do.
BLOG 2026-05-25
The bill finally itemized
Claude Code 2.1.149 splits /usage into skills, subagents, plugins, and per-MCP-server cost. It's a real improvement — and also a concession. The single-number bill outgrew itself, and the receipt now has four columns because the bundle could no longer be sold as one purchase.
BLOG 2026-05-22
The agent that runs while you sleep
Antigravity 2.0 schedules agents to run in the background, and Anthropic's Routines have done it since April. The supervision model we built for AI coding assumed someone was watching each run. Scheduled agents quietly remove the watcher.
BLOG 2026-05-22
The Flash that got expensive
Gemini 3.5 Flash is three times the price of the Flash it replaces, and Google is putting it inside every free product it ships. 'Flash' used to mean cheap. Now it just means fast, and the per-call cost is the part nobody is reading.
BLOG 2026-05-22
You didn't pick this model
On May 17, GitHub made GPT-5.3-Codex the base model for every Copilot Business and Enterprise org. Most developers on those plans didn't choose it and may not be able to change it. The model writing your code is becoming a platform decision.
BLOG 2026-05-20
Parallel agents and the per-developer meter
Cursor 3.5 lets one developer run agents across several repositories at once, and Composer 2.5 makes each call cheaper. Both move the same number in the same direction: how many tokens one person burns in a day.
BLOG 2026-05-20
The 51% number and what "assisted" hides
GitHub reports that around 51% of code committed in early 2026 was AI-generated or assisted. It is the same kind of number as Airbnb's 60%, and it hides the same things — starting with what the word assisted is doing.
BLOG 2026-05-19
The 60% claim, deconstructed
Airbnb's CEO said AI writes 60% of the company's new code. The number is real and the conclusion people draw from it is not. This is the gap between the two.
BLOG 2026-05-10
Why I run an AI chat aggregator alongside Cursor
AI editors like Cursor are great inside the editor. For email drafting, browser-side research, and quick model comparisons, a general AI aggregator does what coding-native tools weren't built for. Here's how I split the two.
BLOG 2026-05-03
AI coding tool pricing is converging — and that's not good news for developers
Cursor, Copilot, Windsurf, and Cline are all landing at $15–20/month per seat. The differentiation is collapsing to model quality and editor integration, and most teams aren't buying for the right reasons.
BLOG 2026-05-03
Context windows stopped being the differentiator and nobody noticed
Two years ago, the context window race mattered. Now every model worth using has 200k+ tokens, and the actual bottleneck has moved somewhere else. Where it moved is more interesting than where it was.
BLOG 2026-05-01
The junior developer question: are AI tools helping or hurting skill development?
AI coding tools handle the kinds of grinding-out-code tasks that used to teach junior engineers their craft. Senior engineers are split on whether this is good. Both sides have a point.
BLOG 2026-04-29
Open-source vs closed-source AI coding tools: the honest tradeoffs
Aider and Cline are open source. Cursor and Copilot aren't. The choice between them comes down to four tradeoffs that aren't usually framed clearly.
BLOG 2026-04-26
Measuring AI coding ROI without fooling yourself
Vendors publish '40% productivity gains'. Internal teams report similar numbers. The reality, when measured carefully, is more modest and more interesting.
BLOG 2026-04-23
The vendor lock-in question for AI editors is bigger than people think
Switching from VSCode to Cursor took an afternoon. Switching back, after a year of accumulated workflow, is a different story. The lock-in isn't where most people are looking.
BLOG 2026-04-15
Why AI coding tools still can't do architecture — and probably won't soon
AI tools handle code-level tasks reasonably well and struggle with architecture-level decisions. The reasons are structural, not just a matter of bigger models.
BLOG 2026-04-12
The '10x faster with AI' claim, deconstructed
Engineers love to claim AI made them 10x faster. They mean something specific by it, and that specific thing isn't what the claim sounds like.
BLOG 2026-04-01
Hiring engineers in 2026: what AI tools change about interviews
If candidates can use AI to solve coding interview questions in 5 minutes, the interview was probably testing the wrong thing. The harder question is what to test instead.
BLOG 2026-03-26
The 7B coding model renaissance: small models are good enough for more than you think
Qwen 2.5 Coder 7B, DeepSeek Coder V2 lite, and a few others are quietly making 'small local model' a real category. The use cases that fit are larger than the 'just for autocomplete' framing.
BLOG 2026-03-23
Code review is now the bottleneck — and AI is making it worse
When AI lets one engineer ship 3x more code, the team's bottleneck moves to review. Most teams haven't adjusted. Here's what's happening and what to do.
BLOG 2026-03-21
The context window arms race ended; nobody noticed
Two years ago, model context windows were the headline feature. Now they range from 128k to 2M and it doesn't matter for most workflows. Here's what actually does.
BLOG 2026-03-19
Why review-as-you-go beats batch review for AI-generated code
Reviewing AI output one chunk at a time feels slower than letting it produce a feature and reviewing the diff at the end. Across many sessions, the reverse turns out to be true.
BLOG 2026-03-19
The most overrated AI coding feature: autonomous debugging
Tools advertise 'AI debugs your code automatically.' In practice, autonomous debugging fails more often than it succeeds, and the failure mode is expensive.
BLOG 2026-03-12
The 'AI doesn't understand my codebase' complaint, examined
Every team using AI tools has someone who's said this. The complaint is sometimes accurate, sometimes a misdiagnosis. Telling the difference matters because the fixes are completely different.
BLOG 2026-02-28
AI tools are forcing a README renaissance, in unexpected ways
Repos with detailed READMEs work better with AI tools. The market is responding. Here's how documentation expectations have shifted.
BLOG 2026-02-12
Why I stopped using AI for naming things
AI tools generate plausible names instantly. The names are almost always worse than what I'd come up with after thirty seconds of thought.
BLOG 2026-02-10
Agentic coding and the end of 'pure' code review
When AI agents produce most of a PR's code, what is review actually for? The answer is shifting in important ways.
BLOG 2026-01-26
The tab completion paradox: simpler features outperform agent loops for daily work
Engineers report Tab autocomplete saves them more time than agent features over a typical day. The marketing emphasizes the wrong thing.
BLOG 2026-01-25
Why agent mode doesn't help me debug, even when it should
Agent loops are designed for autonomous problem-solving. For debugging, the autonomy is precisely what hurts. Investigation requires staying in the loop.
BLOG 2026-01-24
The hidden cost of bad AI tool defaults
Most engineers use their AI tools at default settings. The defaults are often wrong for the engineer's workflow. The friction adds up.
BLOG 2026-01-17
The quiet improvement in AI tool context pruning
Context pruning has gotten dramatically better in 2025. The user-visible effect is subtle. The cost reduction is real.
