Uncontrolled AI Spend Needs a Gateway, Not Another Dashboard
A skeptical guide to testing an open AI gateway, controlling agent spend, and checking caching, telemetry, credentials, and fallback behavior before switching.
TOPIC_INDEX
dev-tools
38 published entries in this topic.
DEV-TOOLS_ENTRIES
RSS← ALL_TOPICSA Four-Step Agent Harness Cut Frontier Usage 75%
A practical local harness routes simple coding work to cheaper models, keeps artifacts auditable, and reserves frontier context for ambiguity.
Coding skills lost 47 points to a short AGENTS.md file
Vercel found a short AGENTS.md docs index reached 100% in its evals, while default skills stayed at 53%. Here is when to use each approach.
128K Context Now Demands Hopper Hardware
FlashPrefill V2 reports 47.26x faster 128K FP8 prefill on H20. Here is the Hopper-only SGLang test path and the benchmark caveat.
Live Code Is Powerful Until Your Agent Can Rewrite It
Autolith v0.35.0 can mutate a live Lisp runtime and recover after crashes. Here is the safe install boundary and the failure test worth running first.
Real-Time TTS Still Breaks at Six Requests Per Second
Nari Labs shows why real-time TTS is a scheduling problem: tuned engines can collapse under load even when one-request latency looks excellent.
Static Checks Miss the PLC Bugs That Matter
SemaPLC shows why AI-generated PLC code needs live runtime tests, not just compilation. Here is the local verification loop worth copying.
Easy OpenClaw Installs Can Expose Your Whole Host
OpenClaw is easy to install, but its Gateway is a trusted control plane. Here is a safer single-user VPS setup with audit commands and clear stop signs.
Safe Rust GPU Code Is Fast Enough to Try Carefully
Rust GPU offload is close enough to CUDA for a serious prototype, but one transfer mistake can make it 400x slower.
The 24GB Local Agent Bet Finally Has a Real Test
Muse Glimmer fits a 24GB local deployment only with headroom discipline. Here is the staged hardware and safety test before giving it tools.
Stop Running Cron on Your Laptop for Claude Code
Claude Code Routines, desktop tasks, and GitHub Actions solve different scheduling problems. Use this decision guide to pick the right execution boundary.
Tool Loops Finally Keep Their Receipts in LLM 0.32
LLM 0.32 adds provider tools, clean CLI output, resumable chains, and inspectable SQLite logs.
DeepSeek Harness Makes Agent Infrastructure Replaceable
DeepSeek Harness makes agent runtimes modular, but its preview status demands a disposable, trace-first test before production credentials touch it.
The AI Adoption Gap Is 8.3x, Not a Model Problem
OpenAI reports an 8.3x enterprise AI usage gap. The fix is not another model, but a staged workflow that turns chat experiments into governed agent work.
Prompt Rules Fail Local Coding Agents Need State Machines
A practical Statewright setup for constraining coding agents by phase, with the 20-line edit cap, 3-file limit, and the limits of its 2/10 to 10/10 report.
Offline Agents Need 14MB, Not a Cloud Bill
Needle 2 fits typed device actions into 14MB. Here is the hardware recipe, benchmark tradeoff, and failure boundary before you put it in a product.
YOLO Coding Needs a MicroVM, Not More Permission Prompts
Docker Sandboxes gives coding agents a private microVM and disposable workspace. Here is a safer unattended setup.
Stop Tuning Coding Agents Without an Eight Task Test
A cheap eight-task probe can tell you whether an AI coding-agent configuration helps before you spend three hours and $150 on a full evaluation.
Chromium Is Too Heavy for Most AI Browsers
Cloudflare Kitesurf makes a useful case for lighter AI browsers, but the deployment choice still comes down to compatibility, session limits, and task shape.
Self Hosting Loses Before Your GPU Pays Back
Self-hosting only wins when utilization is real. Compare optimized API costs with GPU capacity, caching, batch work, and operations before buying hardware.
Voice Agents Need a Live Path and a Slow Path
GPT-Live shows why voice agents need separate clocks for speech and thought. Here is the Realtime API architecture to deploy now.
One AMD GPU Finally Makes DeepSeek V4 Flash Practical
A single AMD MI300X can run DeepSeek V4 Flash, but the real breakthrough is the ROCm work needed to make the awkward hardware behave.
GPU Memory Is the Real Cost of Frontier AI
Cloudflare cut serving costs for Kimi and GLM by shrinking caches and weights, then added integrity checks for the shared memory those gains create.
Tutorials are becoming an agent's missing memory
Microsoft's Resource2Skill turns tutorials, code, and visual references into executable agent skills, scoring 11.9 points above agents without the wiki.
Frontier Coding Just Got Cheaper Than The Hype
Grok 4.5 pairs frontier coding claims with $2 input pricing, 80 TPS, and an open agent harness. The cost story is more convincing than the leaderboard story.
Local AI agents break out of the terminal with Bionic
LM Studio launched Bionic, a desktop agent app for open models that handles coding, documents, and voice — all running locally or with Zero Data Retention cloud routing.
Your AI coding agent burns 40% of context before you type
Systima.ai measured exactly what Claude Code and OpenCode send to the API before your prompt. The 4.7x gap explains why one dashboard climbs while the other stays flat.
That GPT-5.6 migration will cost you more than API credits
A production AI agent company switched from Claude Opus 4.8 to GPT-5.6 Sol and got 2.2x faster at 27% lower cost. The migration took weeks, not minutes.
Your coding agent burns 33k tokens before you type anything
Claude Code sends 33k tokens before you type a word. OpenCode sends 7k. The invisible overhead is eating your context window.
AI agents can finally make your PowerPoint. OfficeCLI makes it happen
OfficeCLI is a single binary that lets AI agents read, edit, and create Word, Excel, and PowerPoint files with visual feedback. 8,300 GitHub stars in six weeks.
GPT-5.5's reasoning gets stuck at 516 tokens. Codex tasks break.
A GitHub issue reveals GPT-5.5 Codex responses cluster at exactly 516 reasoning tokens, correlating with task failures. 390K records analyzed, 82% of exact-516 events come from GPT-5.5.
Free AI model testing on GitHub dies July 30
GitHub Models dies July 30. The free playground, API, and model catalog are all gone. Here's what happened, where to migrate, and why this keeps happening.
Open-weight just became production-ready. Claude Code agrees
ZAI shipped a desktop app that runs Claude Code with GLM-5.2, an open-weight model that matches Opus on benchmarks at a fraction of the cost.
That API call is secretly five models arguing
vLLM Semantic Router turns one API call into a bounded multi-model collaboration. It beats Fugu Ultra on coding benchmarks, and you can run it yourself.
Same resume got 66 to 99. HackerRank's AI hiring is random
HackerRank open-sourced their AI resume scorer. Running the same resume 100 times produced scores from 66 to 99. The code is honest. The scoring is random.
Your AI coding agent is burning money on every cache miss
Reasonix is an open-source terminal agent built specifically around DeepSeek's prefix cache. 99.82% hit rate, $0.05 per turn, and it proves that single-provider agents beat multi-provider ones on cost.
Your AI coding agent wastes tokens on files it doesn't need
Claude Context is an open-source MCP server from Zilliz that gives AI coding agents semantic code search, cutting token usage by ~40%. Hit #1 on GitHub Trending with 10K stars in one week.
Your GPU finally speaks Rust. NVIDIA's compiler is here.
NVIDIA just shipped cuda-oxide, an experimental Rust-to-CUDA compiler that compiles GPU kernels directly to PTX. No DSLs, no C++, just safe Rust on your GPU.