Stable event names plus lint-enforced attributes is the boring discipline that makes logs usable months later. https://blog.sentry.io/structure-a-log/

How to structure a log

A practical guide to writing structured, queryable, lint-enforceable logs with consistent event names, attributes, severity, and schema checks.

blog.sentry.io

Keeping the prompt cache warm during long multi-subagent sessions so you do not pay re-encoding costs when subagents idle is one of those tools that pays for itself immediately. https://github.com/izeigerman/claude-thermos

GitHub - izeigerman/claude-thermos: Keeps your Claude session warm for you

Keeps your Claude session warm for you. Contribute to izeigerman/claude-thermos development by creating an account on GitHub.

github.com

Giving AI agents enforceable WCAG-aligned accessibility rules as a context file is the most direct way to stop them generating inaccessible markup by default. https://github.com/fecarrico/A11Y.md

A file-level tour of agent chat persistence changes how you think about privacy, cleanup, and what delete really means in coding tools. https://x.com/alxfazio/status/2081415650912727439

alex fazio (@alxfazio) on X

the life of a codex conversation on disk

x.com

Cutting most of the system prompt and pushing context into skills, CLAUDE.md and memory is a strong signal that harness design is replacing prompt bloat. https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models

Stopping an agent loop the moment iterations no longer improve output is the simplest possible fix for the runaway-loop failure mode, and open-source changes the economics. https://github.com/loopgain-ai/loopgain

GitHub - loopgain-ai/loopgain: An open-source cost controller for AI agent loops — stops a loop when it's actually converged and rolls back before it degrades, instead of running to a fixed max_iterations cap. Real-time loop-gain (Aβ) bands + best-so-far rollback. Adapters for LangGraph, CrewAI, AutoGen, LangChain, OpenAI Agents, and Claude Agent SDK; raw API for custom stacks.

An open-source cost controller for AI agent loops — stops a loop when it's actually converged and rolls back before it degrades, instead of running to a fixed max_iterations cap. Real-time loop...

github.com

A systematic study of the fake npm package names five LLMs keep inventing, and how quickly attackers register them — the supply-chain risk data the ecosystem needed. https://socket.dev/blog/slopsquatting-targets-across-frontier-llms

New Study Identifies 53 Slopsquatting Targets Across 5 Front...

Five frontier LLMs generated the same nonexistent package names, leaving 53 available for potential slopsquatting across PyPI and npm.

socket.dev

The standout read this week: a rare authentication teardown with concrete trade-offs, from encrypted cookies to S3-backed revocation records at Canva scale. https://www.canva.dev/blog/engineering/session-revocations-at-scale/

Session revocations at scale - Canva Engineering Blog

How Canva keeps hundreds of millions of user sessions fast and secure

canva.dev

Type-aware rules — unhandled promises, unsafe assignments — landing in the Rust linter is the milestone that makes Oxc a genuine full replacement for typescript-eslint. https://oxc.rs/blog/2026-07-22-type-aware-linting-stable

Type-Aware Linting Stable

A collection of high-performance JavaScript tools written in Rust

oxc.rs

An animated explainer on how a petabyte of Postgres across 256 shards presents as a single database — the clearest explanation of distributed query routing I have seen. https://planetscale.com/blog/making-768-servers-look-like-1

Making 768 servers look like 1 — PlanetScale

How to make 768 distinct Postgres servers look like 1 to your applications.

planetscale.com

Claude Code and Codex each catching more bugs in the other model's output than in their own is a routing decision any team running multiple models should act on immediately. https://www.greptile.com/blog/model-inversion

Models are worse at reviewing their own code | Greptile Blog

We noticed GPT reviewed Claude's code better than its own, and vice versa. So we ran a thousand buggy PRs through both models to find out why.

greptile.com

Hiding malware inside SVG images embedded in a fake coding-interview challenge is a specific, reproducible attack that any JavaScript developer could realistically encounter. https://www.elastic.co/security-labs/contagious-interview-malware-svg-steganography

Contagious Interview malware in SVG images: DPRK campaign — Elastic Security Labs

Contagious Interview malware uses SVG steganography to hide payloads in a developer coding challenge. No antivirus vendor detected it.

elastic.co

Data-oriented design producing a JavaScript parser faster than Oxc, SWC, and Babel — with the engineering behind it explained step by step. https://www.arshad.fyi/writings/engineering-high-performance-parsers

Engineering High-Performance Parsers with Data-Oriented Design

Notes from building Yuku: the AST is flat arrays of u32 indices instead of a pointer tree, and memory layout, allocation, strings, unicode, and serialization all follow from that one decision.

arshad.fyi

Designing for the case where an agent edits the same document you are is a UX problem nobody has a good answer to yet — Sunil Pai frames it sharply. https://sunilpai.dev/posts/one-document-two-hands

one document, two hands

the agent belongs beside you, not between you and the app

sunilpai.dev

AI tools now write Lean proofs for results they found themselves — the formalization research feedback loop has inverted and human mathematicians are racing to keep up. https://xenaproject.wordpress.com/2026/07/20/human-mathematicians-are-being-outcounterexampled/

Human mathematicians are being outcounterexampled

It’s been an interesting few weeks for counterexamples. This post is basically my perspective of what has been going on in the world of formalization, AI tools and, in particular, counterexam…

xenaproject.wordpress.com

Fusing one model's architecture into the circuitry instead of running any model on general hardware is the most structurally different AI hardware bet announced this year. https://thenextweb.com/news/google-frozen-chip-gemini-silicon

Having the frontier model dump its understanding to a todo list then handing execution to a cheaper model is a concrete, adoptable pattern that cuts cost without sacrificing quality. https://stencil.so/blog/prewalk

You only need the frontier model for one single edit

Hand the cheap model a trajectory, not a plan: /prewalk keeps 92–97% of the frontier pass rate at roughly half the cost, measured on SWE-Bench Pro.

stencil.so

Cursor's breakdown of how planner-plus-worker swarms beat a single frontier model on speed, quality, and cost is the clearest case so far for moving beyond the single-model loop. https://cursor.com/blog/agent-swarm-model-economics

Agent swarms and the new model economics · Cursor

We compared old and new agent swarms building SQLite from scratch and found that better coordination delivers similar quality at a fraction of the cost.

cursor.com

Structuring agent tasks into deterministic code to cut token usage 8x is one of the most concrete cost-reduction stories of the year, with real before-and-after numbers. https://www.adamhjk.com/blog/a-practical-guide-to-reducing-token-spend/

A Practical Guide to Reducing Token Spend · Adam Jacob

Building a Swamp Workflow reduced real world token usage of an AI Agent by 8x and improved execution time by 2x - here is how to do it yourself.

adamhjk.com

Matteo Collina's definitive answer: you cannot harden Node core against prototype pollution, and the fix belongs at the boundary — not in the runtime. https://adventures.nodeland.dev/archive/no-we-cant-harden-nodejs-against-prototype/

No, We Can't Harden Node.js Against Prototype Pollution

"Prototype pollution isn't a Node or library issue, it's on the app's guard to thwart code vulnerabilities!"

adventures.nodeland.dev

Fetch collapsing every network failure into a bare TypeError is one of the most frustrating gaps in the web platform — this TC39 proposal threads typed error codes through the stack and would actually fix it. https://jasnell.me/posts/fetch-needs-error-codes

Intercepting dangerous git and shell commands before an agent runs them sits exactly between trusting the agent and reviewing every diff — the missing piece for unattended agent workflows. https://github.com/Dicklesworthstone/destructive_command_guard

GitHub - Dicklesworthstone/destructive_command_guard: The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents.

The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents. - Dicklesworthstone/destructive_command_guard

github.com

How Netflix benchmarked vLLM and Triton then built their own serving stack instead of paying for hosted APIs is a proper architecture teardown. https://netflixtechblog.com/in-house-llm-serving-at-netflix-a5a8e799ea2c

In-House LLM Serving at Netflix

By AI Platform’s Model Runtime team and Inference team

netflixtechblog.com

Disposable VMs for coding agents behind a per-session network allow-list means an agent can install, build, and test without touching the host or reaching any credential it should not. https://github.com/clawkwork/clawk

GitHub - clawkwork/clawk: Give coding agents a disposable Linux VM, not your laptop

Give coding agents a disposable Linux VM, not your laptop - clawkwork/clawk

github.com

Replaying a coding agent's session on a 3D citymap of your repo makes the traversal pattern, dead ends, and repeated returns obvious in a way no log file ever could. https://github.com/cosmtrek/mindwalk

GitHub - cosmtrek/mindwalk: A visualization tool that replays coding-agent sessions on a 3D map of your codebase.

A visualization tool that replays coding-agent sessions on a 3D map of your codebase. - cosmtrek/mindwalk

github.com

A useless if that never changes the outcome makes a latency-bound loop four times faster by giving the CPU a predictable branch — the kind of micro-insight you cannot unsee once you know it. https://purplesyringa.moe/blog/quadrupling-code-performance-with-a-useless-if/

The person who coined context engineering on finding the dumb zone, intentional compaction, and why you-are-completely-right means start a new session. https://newsletter.pragmaticengineer.com/p/context-engineering-with-dex-horthy

Context engineering with Dex Horthy

Dex Horthy explains why context engineering is key to building more effective AI-assisted software without sacrificing code quality.

newsletter.pragmaticengineer.com

Using CHECK constraints to correlate a non-partition column with the partition key — sequential session IDs in time-partitioned data — buys pruning you would otherwise write off as impossible. https://hakibenita.com/postgresql-partition-pruning

How to Achieve Pruning When Querying by Non-Partitioned Columns in PostgreSQL

hakibenita.com

A flagship model documented to take destructive actions — including deleting files — as long as they are not unambiguously prohibited is a reminder that safety evaluations lag capability releases.

OpenAI's new flagship model deletes files on its own, people keep warning | TechCrunch

A number of social media posts claim that GPT-5.6 Sol deleted files and data without warning. OpenAI had basically disclosed the problem in June.

techcrunch.com

Auto-executing malicious binaries on project open with no user interaction, still unfixed after seven months of disclosure, is the attack surface that agent-heavy IDE workflows can least afford. https://mindgard.ai/blog/cursor-0day-when-full-disclosure-becomes-the-only-protection-left