Mermaid to Unicode box art (grok-mermaid)
Willison extracts a self-contained Rust terminal renderer for Mermaid diagrams from the newly open-sourced Grok CLI, turning flowcharts into Unicode box art you can run standalone.
xai-org/grok-build, now open source
A close read of xAI's newly open-sourced grok CLI, including the community backlash over it uploading entire working directories to xAI's cloud buckets.
How I tricked Claude into leaking your deepest, darkest secrets
Willison highlights Ayush Paul's discovery of a data-exfiltration hole in Claude's otherwise carefully-designed web_fetch tool, a sharp lesson for anyone building tool-using agents.
Dependabot's new three-day release cooldown
A note on GitHub making Dependabot wait three days after a release before opening update PRs — a sensible default guard against compromised or yanked packages.
simonw/pedalican
Willison stumbles onto Codex Desktop's animated 'pet' feature — a Clippy-like robot you can customize — and builds his own, a small, playful look at where coding tools are going.
Voices · last 30 days
From engineers worth following
Latest posts from a curated roster — infra, devops, AI, backend.
Context engineering with Dex Horthy
Dex Horthy explains why deliberately engineering the context you feed AI tools — not just prompting — is key to getting effective, high-quality code out of them.
What is 'loop engineering?'
Orosz investigates the buzzy term 'loop engineering' — triggers, cron jobs, and agentic loops — and asks whether it's a durable practice or a here-today-gone-tomorrow trend.
The Pulse: What can we learn from Bun's rapid Rust rewrite with AI?
A breakdown of Bun's 11-day, $165K-in-tokens Rust rewrite and what it reveals about the real state of agentic coding, plus the heating-up coding-LLM wars.
The Pragmatic Engineer AMA
Orosz answers listener questions on AI, engineering, hiring, and careers in an AMA episode.
Tech jobs market in 2026, part 3: hiring managers & job seekers
Drawing on 50+ hiring managers and job seekers, a look at a strange tech market: nobody finds each other, AI roles are hottest, and it's tough for engineering leaders.
DSLs Enable Reliable Use of LLMs
Unmesh Joshi argues that domain-specific languages give LLMs the clear boundaries they need to generate exactly what's intended, acting as a harness from the very start.
Fragments: July 13
Notes from Thoughtworks' Future of Software Development retreat, including a whole session on the newly-named practice of 'Harness Engineering.'
Experiences with local models for coding
Birgitta Böckeler reports on actually using local LLMs for coding, comparing them on standard tasks and trying the most promising one day to day.
Viability of local models for coding
A companion memo laying out the factors — hardware, model size, task type — that determine whether local LLMs are viable for real programming work.
Fragments: July 6
More fragmentary thoughts from the European edition of Thoughtworks' Future of Software Development retreat, spanning five parallel streams.
Own the Outer Loop
Osmani argues that 'loop engineering' still needs a human at the boundary — owning the outer loop of judgment and direction while agents handle the inner work.
Agentic Autonomy Levels
A working model for thinking about degrees of agent autonomy, giving teams a shared vocabulary for how much independence to grant coding agents.
Agentic Code Review
Osmani makes the case that code review is now the most leveraged skill in software, as reviewing agent output becomes the core human bottleneck.
The one thing every experienced engineer gets wrong
Teller observes that even strong engineers stumble on project management for their first big project, making it a rite of passage worth naming.
Theory of constraints, AI, and code review
Applying theory of constraints to AI-assisted work: expert feedback is how you learn taste, and your code doesn't matter until it's actually shipped.
Using vibes to rewrite my blog
A hands-on account of vibe-coding a blog rewrite, arguing the death of engineering is exaggerated — you still need to know what you're doing.
You get what you hire
Teller reflects that hiring engineers is really deciding what kind of system you'll have — obvious in theory, surprising to watch play out firsthand.
Code is the easy part
A startup war story about how fixing one janky script meant refactoring half the business — a reminder that code is rarely the hard part.
Generated and suppressed demand
Larson revisits his framework for struggling teams, examining how demand gets both generated and suppressed and what that means for staffing and backlogs.
Make no assumptions
Using the geology metaphor of soil horizons, Larson explores the layered assumptions beneath organizations and why surfacing them matters for leaders.
Have you heard? Clickhouse is winning the observability wars!
Majors revisits her columnar-storage prediction for observability, and skewers vendors who build it, nerf it, and resell a worse version as 'Datadog but cheaper.'
In defense of AI mandates
A contrarian argument that when you need a coordinated change on a tight timeline, a top-down mandate is the most honest and effective way to fund it.
Is it ethical to use AI?
Majors on externalities, engagement, and the trap of purity politics, arguing the healthiest move is to make AI boring again.
Meet Alice. Alice is impatient.
Brooker uses an impatient user named Alice to explain why averages mislead when reasoning about latency, and why tail behavior is what your users actually feel.
Is this blog written by AI?
Brooker flatly states none of his blog's prose is AI-written — the quirks, typos and em-dashes are all his — a small stand for human voice in the LLM era.
Do Automated Evals Work?
Husain compares 100 human-annotated traces against automated eval systems to test, empirically, how much you can trust automated evaluation of AI output.
'It's Hard to Eval' Is a Product Smell
Drawing on three years focused on evals, Husain argues that products which feel hard to evaluate are usually hard for users to verify too — a design smell.
6 months to live for open models
Lambert argues the most serious test yet of open-source AI's viability is happening right now, and lays out what determines whether open models survive.
Latest open artifacts (#22): Zyphra, Cohere, and Poolside
His regular survey of new open model releases, assessing how Zyphra, Cohere and Poolside are broadening the ecosystem and why they're releasing weights.
GLM-5.2 is the step change for open agents
Lambert makes the case that GLM-5.2 crosses a capability threshold he'd been watching, marking a genuine step change for open-weight agents.
Banning Open Source AI Would Be A Mistake
An op-ed, co-authored with Kevin Xu, arguing for a general audience why restricting open-source AI would be a strategic and practical error.
State of the blog, mid-2026
A reflective check-in roughly three years into writing Interconnects weekly, on where the blog and the open-model conversation stand.
A broken DNSSEC rollover took down .al
When a failed DNSSEC key rollover broke the .al TLD, Cloudflare deployed a Negative Trust Anchor and now uses the new EDE 33 error code so 1.1.1.1 clients can see when validation was bypassed.
Introducing Precursor: detecting agentic behavior with client-side signals
Cloudflare's new bot-management engine turns continuous session-level behavior into detection signals, giving sites visibility into how humans and bots actually interact.
Improving Smart Tiered Cache for public cloud regions
Smart Tiered Cache now supports precise upper-tier selection for origins on AWS, GCP, Azure and Oracle Cloud using customer-provided region hints.
Why we cannot wait for better post-quantum signature algorithms
Surveying NIST's nine candidate post-quantum signature algorithms, Cloudflare argues that despite their promise, we should deploy ML-DSA now rather than wait.
Introducing Meerkat: an experiment in global consensus
Cloudflare Research unveils Meerkat, a global consensus service built on a new algorithm called QuePaxa, aimed at a strongly consistent, fault-tolerant key-value store.