DEV Community: Rohit Raj

DeepSeek Harness vs Claude Code vs Codex CLI: The v0.1 Developer Preview, Honestly — 2026

Rohit Raj — Fri, 14 Aug 2026 03:17:50 +0000

DeepSeek open-sourced its agent harness on August 13, 2026 under MIT — 572 points on Hacker News in a day. Every page currently ranking for "DeepSeek Harness vs Claude Code" describes a different, older product. Here is what v0.1 actually ships: the Cordis plugin kernel, the real cordis.yml keys, 53 built-in tools, and the one config default that decides whether it touches your repo.

Read the full version with code samples, diagrams, and architecture details: DeepSeek Harness vs Claude Code vs Codex CLI: The v0.1 Developer Preview, Honestly — 2026

More engineering notes: rohitraj.tech/en/notes

This Week in AI Dev: Models Day, Stolen Reasoning, and a $16,000 GPU (Week 33 of 2026)

Rohit Raj — Thu, 13 Aug 2026 09:46:51 +0000

Week 33 of 2026 put three frontier models on the table inside 48 hours — DeepSeek V4 Pro at 1.6T params under MIT, Grok 4.6 one point behind Claude Opus 5, Meta back in open weights with Muse-Glimmer-30B. Then a paper showed the encrypted reasoning those closed models hide can be decoded by a weaker sibling model, and 315,320 scraped blocks gave up 182 credentials.

Read the full version with code samples, diagrams, and architecture details: This Week in AI Dev: Models Day, Stolen Reasoning, and a $16,000 GPU (Week 33 of 2026)

More engineering notes: rohitraj.tech/en/notes

Prime Agent: The RLM + Continual Harness Guide (And When to Skip It) — 2026

Rohit Raj — Sun, 09 Aug 2026 06:37:49 +0000

Prime Intellect open-sourced Prime Agent on August 5, 2026 and it hit #1 on GitHub trending with +2,483 stars in 24 hours. It scores 95.5% on ARC-AGI-3 with Opus 5 — above the 95.4% human expert baseline. Here's what a Recursive Language Model actually is, the code that makes it different, and the line in the README that should stop you shipping it to production this week.

Read the full version with code samples, diagrams, and architecture details: Prime Agent: The RLM + Continual Harness Guide (And When to Skip It) — 2026

More engineering notes: rohitraj.tech/en/notes

Cloudflare Computer vs Cloudflare Sandbox: Which Agent Runtime Should You Actually Use? (2026)

Rohit Raj — Fri, 07 Aug 2026 06:16:14 +0000

Cloudflare shipped @cloudflare/computer on August 3 and it hit #1 on GitHub trending with 2,802 stars in a day — four months after Sandboxes went GA. Cloudflare's own docs never compare the two, so here's the decision table, working wrangler.jsonc code, and the preview-status caveat the launch coverage buries.

Read the full version with code samples, diagrams, and architecture details: Cloudflare Computer vs Cloudflare Sandbox: Which Agent Runtime Should You Actually Use? (2026)

More engineering notes: rohitraj.tech/en/notes

TencentDB Agent Memory v2.0 Review: A Self-Hosted Team Memory Hub for AI Agents (2026)

Rohit Raj — Thu, 06 Aug 2026 05:27:21 +0000

Tencent's Agent Memory hit #1 on GitHub trending this week at 15.3k stars — a self-hosted hub that turns chats, docs, and code into four governed memory assets shared across Claude Code, OpenClaw, and Hermes. I hand-roll this exact stack daily, so here's what the productized version gets right, the vs-Mem0/Zep/Letta comparison nobody else wrote, and the governance wiring you need before your team touches it.

Read the full version with code samples, diagrams, and architecture details: TencentDB Agent Memory v2.0 Review: A Self-Hosted Team Memory Hub for AI Agents (2026)

More engineering notes: rohitraj.tech/en/notes

Shieldstral vs Llama Guard vs OpenAI Moderation API: A Self-Hosted Content Moderation Guide (2026)

Rohit Raj — Wed, 05 Aug 2026 08:31:07 +0000

Mistral released Shieldstral on August 4, 2026 — a 3B open-weights safety classifier that reads your moderation policy at inference time instead of training on fixed categories. Here's how it actually compares to Llama Guard and the OpenAI Moderation API, the real self-host commands, and the one case where I'd still reach for a fixed taxonomy instead.

Read the full version with code samples, diagrams, and architecture details: Shieldstral vs Llama Guard vs OpenAI Moderation API: A Self-Hosted Content Moderation Guide (2026)

More engineering notes: rohitraj.tech/en/notes

This Week in AI Dev: Everything Got Smaller, and Six CVEs Turned Out to Be Fake (Week 32 of 2026)

Rohit Raj — Tue, 04 Aug 2026 05:50:02 +0000

Week 32 of 2026 was a compression week. MiniMax-H3 shipped open weights that fall from 123.6 GB to 42.5 GB, Cloudflare doubled Kimi K2.6 context to 1.37M tokens with an FP8 KV cache, and AirLLM hit 27.5k stars streaming a 2.8T model through 4 GB of VRAM. Meanwhile JFrog found 54 of 55 SQLite CVEs from one repo were AI-generated fiction — one briefly scored a CVSS 10.0.

Read the full version with code samples, diagrams, and architecture details: This Week in AI Dev: Everything Got Smaller, and Six CVEs Turned Out to Be Fake (Week 32 of 2026)

More engineering notes: rohitraj.tech/en/notes

DeepSeek DSpark in llama.cpp: How to Get 2x Local Inference on V4-Flash-0731 (2026)

Rohit Raj — Mon, 03 Aug 2026 03:28:28 +0000

llama.cpp merged DeepSeek V4 DSpark support on August 2, 2026 — the docs still say Qwen3-only. Here are the actual flags, the measured 39.95 to 79.93 tokens/sec jump, why the config with the higher acceptance rate is the slower one, and the RAM you need before any of it matters.

Read the full version with code samples, diagrams, and architecture details: DeepSeek DSpark in llama.cpp: How to Get 2x Local Inference on V4-Flash-0731 (2026)

More engineering notes: rohitraj.tech/en/notes

This Week in AI Dev: Agent Skills Went GA, Then the Benchmark Landed (Week 31 of 2026)

Rohit Raj — Thu, 30 Jul 2026 04:43:22 +0000

Week 31 of 2026 shipped one story twice. GitHub made agent skills and MCP generally available in Copilot code review, Google added hooks to Gemini Managed Agents — and on the same day, a benchmark showed the best model follows a written policy document just 36.2% of the time, while a self-propagating worm walked through Copilot for Word. Plus what Kimi K3 really costs to run locally (594 GB floor) and the 250B sleeper release you can actually serve.

Read the full version with code samples, diagrams, and architecture details: This Week in AI Dev: Agent Skills Went GA, Then the Benchmark Landed (Week 31 of 2026)

More engineering notes: rohitraj.tech/en/notes

OpenAI Codex Security vs Snyk vs Semgrep vs CodeQL: What the New Open-Source Scanner Actually Changes (2026)

Rohit Raj — Wed, 29 Jul 2026 14:52:29 +0000

OpenAI open-sourced Codex Security under Apache-2.0 — a CLI and TypeScript SDK that builds a threat model of your repo, then runs an isolated validator to prove a finding is exploitable before it reports it. Here's how that differs from Snyk, Semgrep and CodeQL, the real commands, and the three cases where I'd still reach for Semgrep instead.

Read the full version with code samples, diagrams, and architecture details: OpenAI Codex Security vs Snyk vs Semgrep vs CodeQL: What the New Open-Source Scanner Actually Changes (2026)

More engineering notes: rohitraj.tech/en/notes

Block's Buzz (2026 Guide): Self-Host the Workspace Where AI Agents Are Teammates, Not Bots

Rohit Raj — Fri, 24 Jul 2026 06:38:34 +0000

Block released Buzz on July 21, 2026 — an Apache-2.0, self-hostable workspace built on Nostr where AI agents join channels as cryptographically-signed members, not permission-restricted bots. It hit 7,600+ GitHub stars in three days. The launch coverage tells you what it is; this guide shows you how to actually run it: the exact install path, how to onboard a Claude Code or Codex agent with its own keypair, where Buzz genuinely beats Slack-plus-bots, the compliance gaps that should keep it out of production today, and the hardening checklist the same week's OpenAI–Hugging Face incident makes non-negotiable.

Read the full version with code samples, diagrams, and architecture details: Block's Buzz (2026 Guide): Self-Host the Workspace Where AI Agents Are Teammates, Not Bots

More engineering notes: rohitraj.tech/en/notes

Gemini 3.6 Flash vs 3.5 Flash-Lite: Which One to Ship — and the Price Hike Nobody Leads With (2026)

Rohit Raj — Thu, 23 Jul 2026 11:43:35 +0000

Google's July 21 drop is an efficiency release, not an intelligence release: Gemini 3.6 Flash scores the same Intelligence Index as 3.5 Flash but finishes tasks in half the time at a lower per-task cost — while Flash-Lite quietly got a 67% output-price increase. Here's the real per-task math, the migration code, and the tier decision I'd actually ship.

Read the full version with code samples, diagrams, and architecture details: Gemini 3.6 Flash vs 3.5 Flash-Lite: Which One to Ship — and the Price Hike Nobody Leads With (2026)

More engineering notes: rohitraj.tech/en/notes