AI Articles & Guides

Deep dives into local LLMs, Model Context Protocol (MCP), autonomous agents, and prompt engineering architectures.

Showing 112 of 126 articles

Cross-Platform Agent Skills: Writing Universal SKILL.md for Antigravity, OpenCode & Claude Code

Cross-Platform Agent Skills: Writing Universal SKILL.md for Antigravity, OpenCode & Claude Code#agent-skills

A complete guide to authoring portable Agent Skills that run seamlessly across Google Antigravity, OpenCode, Claude Code, and Cursor without rewrite.

Aug 18, 2026

TH

Malicious Agent Skills & Supply Chain Attacks: Securing the SKILL.md Ecosystem

Malicious Agent Skills & Supply Chain Attacks: Securing the SKILL.md Ecosystem#agent-skills

How third-party AI agent skills introduce hidden RCE, environment variable exfiltration, and prompt injection — and how to sandbox and audit SKILL.md packages.

Aug 18, 2026

TH

Cloud GPU Benchmark: Hosting Qwen 3.8 & DeepSeek on RunPod vs Together AI

Cloud GPU Benchmark: Hosting Qwen 3.8 & DeepSeek on RunPod vs Together AI#runpod

Cost, latency, and throughput comparison for hosting Qwen 3.8 and DeepSeek-Coder on RunPod spot GPUs versus Together AI serverless inference.

Aug 18, 2026

IN

Catastrophic Remembering: Why Your CLAUDE.md Keeps Growing

Catastrophic Remembering: Why Your CLAUDE.md Keeps Growing#agentic-prompts

Agentic prompt files like CLAUDE.md and AGENTS.md grow +226% over their lifetime. Latent-reasoning comments halt the growth and lift instruction-following.

Aug 17, 2026

TH

Coding Agents at Production Scale: What 761M LLM Calls Reveal

Coding Agents at Production Scale: What 761M LLM Calls Reveal#copilot

A production-scale study of Copilot's agent: 761M calls, 95T tokens. KV cache collapses across turns and after model switches; failures trigger 4x compute.

Aug 17, 2026

TH

DeepSeek Harness: Everything Is a Plugin

DeepSeek Harness: Everything Is a Plugin#deepseek

DeepSeek Harness (dsh) is DeepSeek's MIT-licensed agent runtime where everything is a plugin — the four presets, setup, and what it means for benchmark trust.

Aug 17, 2026

TH

Frontier Agent Benchmarks: DeepSeek V4 vs Qwen 3.8

Frontier Agent Benchmarks: DeepSeek V4 vs Qwen 3.8#deepseek

DeepSeek V4 vs Qwen3.8-Max across agentic benchmarks: verified scores, vendor-reported caveats, pricing, and which model fits your workload.

Aug 17, 2026

TH

Gemini 3.7 Flash: The Workhorse Gets Smarter and Cheaper

Gemini 3.7 Flash: The Workhorse Gets Smarter and Cheaper#gemini

Gemini 3.7 Flash ships three weeks after 3.6 Flash with big coding and agent gains at half the price. FrontierCode, DeepSWE, AutomationBench, and what changed.

Aug 17, 2026

TH

Harness-IF: Are Coding Agents Following Rules, or Just Doing What They'd Do Anyway?

Harness-IF: Are Coding Agents Following Rules, or Just Doing What They'd Do Anyway?#harness-if

A benchmark that separates compliance from coincidence: every model is worse at rules opposing its defaults — and where you put the rule changes everything.

Aug 17, 2026

TH

Qwen3.8-27B: Local Coding Agents, and the Overthinking Problem

Qwen3.8-27B: Local Coding Agents, and the Overthinking Problem#qwen

Qwen3.8-27B is an Apache 2.0 local agent model that beats Claude Opus 4.8 on agentic tasks — but its default reasoning over-thinks. Setup, MTP, and the right settings.

Aug 17, 2026

TH

SWE-RPG: Coding Agents Fail at Implicit Requirements, Not Code

SWE-RPG: Coding Agents Fail at Implicit Requirements, Not Code#swe-rpg

Agents resolve only 31.5% of real issues — and implicit requirement recovery is the biggest bottleneck (24.5–46%). The fix is asking clarifying questions.

Aug 17, 2026

TH

SWE-Touch: Coding Agents Break When Users Touch the Code

SWE-Touch: Coding Agents Break When Users Touch the Code#swe-touch

Coding agents drop 7.7 points on SWE-bench Verified when users edit code mid-task. They don't re-inspect the workspace, and open models degrade most.

Aug 17, 2026

TH
...