Reviews of AI coding assistants, agentic tools, and the workflows that make them useful.
42 articles
xAI's own model card reports Terminal-Bench 3.0 at 26.0%. Artificial Analysis reports 88.4% on 2.1. Both are the same model. The model card is unusually honest about why — the problem is everyone quoting it drops the qualifiers.
No DeepSeek-V4 config is public yet. The V3 one is, and its KV-cache math tells you what a million-token window actually costs in GPU memory.
A year of one scheduled digest agent in production: token costs per stage, infrastructure line items, and the 41 runs that needed a human.
Four failure classes that survive a clean parse, and the three-layer validation pass we run instead.
A per-channel ledger makes cross-posting resumable. Why exit code 0 is not a receipt, and what to do when an API has no idempotency key.
Three credential models for non-human callers: static keys, the OAuth 2.0 device grant, and short-lived scoped tokens - and when each one fits.
A field-by-field guide to API error bodies: stable codes, retryable flags, wait hints, fix examples, and the shapes that trap agents in retry loops.
Ground truth files, an interface contract, one acceptance command, and explicit out-of-bounds rules -- so the agent runs end to end without babysitting.
It runs as a child process with your user's permissions. Four commands to see which files it opens and hosts it dials, plus three containment changes.
A mechanism-level comparison of three AI pull request reviewers — how each one builds context, how noisy it is by design, and how to bake them off on your own repo before buying seats.
OpenCode and Claude Code bill by the token, so cost per task is the metric that matters. Here is how we tracked it and what we learned.
How we split work across OpenCode, Claude Code, and Aider in one terminal without losing track of which agent changed what.
Configure providers, context files, and project conventions so the agent produces useful output on an existing codebase from day one.
We ran OpenCode against a local Ollama model on a proprietary codebase, with no source code sent to a cloud API.
We ran it as a daily driver for two weeks. It edits files, runs tests, and supports multiple models without replacing your IDE.
A hands-on look at two open-source assistants: how each handles model choice, repo context, and your git history.
A practical look at which Model Context Protocol servers actually earn a slot in your editor config, what they do, and where they break down.
Compared on codebase context, review depth, and noise, plus which one fits the way your team actually merges pull requests.
How to split one large refactor across subagents: scoping each task, isolating file conflicts, and reviewing the merged result.
Roo Code began as a Cline fork. Here is how the two open-source, bring-your-own-key VS Code extensions actually differ.
I tested Aider on 9 projects with 6 LLMs over six weeks for $47.30. Why git-native beats accept/reject buttons, and where terminal-only falls short.
v0 generates React components with shadcn/ui, Tailwind, and TypeScript. Here is how they held up under real product requirements.
Orthrus injects diffusion attention into every layer of a frozen autoregressive Transformer, leaving the base model's output distribution unchanged.
Warp compiles Python to CUDA kernels. We benchmarked it against JAX and Taichi to find where it earns a spot in your stack.
Macchiato's day-2 build adds a live token/cost sidebar and keyboard shortcuts for swapping between Claude Code and OpenCode in one terminal. Here's what shipped and what it means.
We benchmarked DeepSeek Coder, Qwen 2.5 Coder, and CodeLlama on Apple Silicon and NVIDIA GPUs for latency, code accuracy, and readiness to replace cloud APIs.
A plugin for VS Code and JetBrains with customizable context and a transparent architecture. Where it replaces Copilot, and where it does not.
v0 generates React and Next.js UI from natural language prompts. A pragmatic look at what it produces, how the output compares to hand-written code, and when it saves real development time.
GitHub's standalone Copilot desktop app changes the matchup. We compare workflow surface, approval semantics, and model neutrality.
Searle's Chinese Room, stochastic parrots, and IIT predict where LLMs break -- and what that means for prompts, retrieval, and agent loops.
Running Llama 3.3 70B on an M-series Mac Studio versus paying per token: here's the math at typical developer volumes, and three cases where local still wins.
We tested Bun 1.2 against Node.js 22 LTS on real workloads: where the speed gap is real, where Node compatibility breaks, and how to decide on migrating.
Nous Research's tool installs with one shell command and keeps agent memory in local files. We compare its file-based approach to Mem0 and Letta.
We tested both on a Next.js app, a Python CLI, and a Rust library migration. Cursor won on velocity -- but one scenario still favors Copilot.
The coding agent runs inside a browser tab. Which workflow patterns pay off, the limits, and how it compares to Codex CLI and IDE agents.
We ran both on the same codebase across refactoring, debugging, and agentic tasks. What each shipped, and the speed-vs-cost tradeoff.
How multi-agent review catches what single-pass LLM reviews miss, and where AdamsReview fits in your pipeline.
Replaced youtube-dl for programmatic video and audio extraction: install, format selectors, the Python API, and gotchas we hit across three real workflows.
The build-your-own-x GitHub repo has 350k+ stars. These 10 picks cover databases, compilers, Git, and neural nets, all from scratch.
A measured look at where this category fits, which workflows benefit, and what to verify before you switch.
Daniel Stenberg confirmed the bug in one of the most-reviewed codebases on the planet. What it means for AI-assisted security review.
Model size math, real tokens/sec for 7B-32B models, and when to reach for Ollama, llama.cpp, or MLX.
One email every Friday with the tools worth your time. No spam, unsubscribe anytime.