AI Agents

Reliability, delivery control and real-world adoption of AI agents: how to make an agent prove the work is done before it touches production.

Articles

The Language of Your Prompt Decides Which Parallel Universe of Training Data You Are Talking To 2026/10/15
Stop translating your Prompt into English to save Token! Why does the more you translate into English, the more the AI answers with "給到" and "質量"? Drawing on ACL language model research and hands-on experience, this article explains why "technical terms in the original + native-language description + explicit regional anchors" is the best way to cut the cost of back-and-forth corrections.
Let AI Be Your Mentor: A Self-Review Routine with Claude Code /insights and Codex Computer History 2026/10/13
If you use both Claude and Codex and have Computer History turned on, two small tricks go a long way: ask Codex to act as your mentor and review your recent work using Computer History, then run /insights in Claude Code to get a review report that moves from the big picture to the small details, so you end up with a summary for steady growth.
Introducing Deckhand: Turn Claude Code into Your Multi-Agent Mission Control 2026/10/11
Deckhand is an open-source Claude Code plugin that adds a toolbar above the input box: dispatch tasks read-only to Codex, Cursor agent, and agy for a second opinion, run parallel subagents with Sub5, and track usage metrics while switching models. This post covers installation, features, data flow, and known limitations.
Stop Mindless Fan-out: A Practical 4-Layer Orchestration Architecture with Claude Haiku 5.5 2026/10/08
Claude Haiku 5.5 is fast and cheap, but throwing every task at it in parallel only amplifies misjudgments, reviewer bias, and Supply Chain risks. This article breaks down a layered architecture—Fable for judgment, Opus for orchestration, Sonnet and Haiku for implementation, CI for verification, and Opus for acceptance—and explores which tasks actually warrant a Fan-out.
Stop Dumping SOPs Directly into AI Agents: Precisely Control Agent Behavior with RFC 2119 (MUST/SHOULD/MAY) 2026/10/08
Dumping human SOPs directly into an LLM often leaves agents succeeding once only to fail three times. This article demonstrates how to rewrite System Prompts and Agent Skills using RFC 2119's MUST, SHOULD, and MAY—clearly drawing the line between hard constraints, strong recommendations, and autonomous discretion—complete with three practical implementation principles.
An AI Agent Is Basically an Agile Team: Demystifying 5 Core Concepts Through Everyday Software Engineering 2026/10/08
Breaking down the 5 core concepts of an AI Agent in plain software engineering terms: Prompt, Context, Harness, Loop, and Graph—mapped directly to ticket titles, specs, dev environments & guardrails, test-and-fix loops, and a PM's workflow chart.
Making AI Answers More Readable: Karpathy on ASD-STE100 2026/10/05
Karpathy recommends asking AI to explain things using ASD-STE100 to reduce misinterpretation and ambiguity. Here is a production-ready template to drop into your global CLAUDE.md or AGENTS.md.

Notes

Notes · 2026/10/15 11:07

Which parallel universe of the AI world are you talking to?

Read the full article：The Language of Your Prompt Decides Which Parallel Universe of Training Data You Are Talking To
Notes · 2026/10/12 11:33

Cursor Pro at $20/mo: Is It Worth It? Huge Value Deal

A single Cursor Pro subscription gets you:
- Simultaneous access to both the IDE editor and the Agents Window.
- Grok Bot login with the same account, plus a custom email address (yourname@mail.grokbot.com).
- Access to multiple AI models: Grok, Claude Fable, Gemini, Muse Spark, and GPT-5.6 Sol.
- Syncing GitHub repos to Cursor Origin—Cursor's code hosting platform built for AI agents, currently in early beta.
For casual AI users, this is an absolute steal.

For power users, third-party model limits mean using Cursor as a sidekick alongside your primary tool, but the overall payoff is still well worth it.
Notes · 2026/10/12 00:11

Deckhand・Turn Claude into Your Multi-Agent Mission Control

(Custom Claude Code Mods)

Adds a toolbar right above the input box: monitor usage, switch models in a click, run parallel subagents with Sub5, and call Codex, Cursor agent, and agy read-only for second opinions.

Claude stays in charge as the primary agent; all other AIs are strictly read-only.

Installs in two commands, MIT licensed.

Read the full article：Introducing Deckhand: Turn Claude Code into Your Multi-Agent Mission Control
Notes · 2026/10/08 14:06

[Claude] Building a low-cost, high-defense Agent architecture: from Fable decomposition and Opus orchestration to Sonnet implementation and isolated Worktree verification

Read the full article：Stop Mindless Fan-out: A Practical 4-Layer Orchestration Architecture with Claude Haiku 5.5
Notes · 2026/10/08 13:18

Dumping an SOP straight into an AI Agent often means it runs right once, then fails three times.

The problem usually isn't the model—it's that your rules lack clear enforcement levels.

Rewrite your SKILL using RFC 2119's MUST/SHOULD/MAY:
- MUST: Never skip under any circumstances
- SHOULD: Follow by default; justify any deviation
- MAY: Leave entirely to model discretion
That is how you craft a SKILL that actually delivers in production.

Read the full article：Stop Dumping SOPs Directly into AI Agents: Precisely Control Agent Behavior with RFC 2119 (MUST/SHOULD/MAY)
Notes · 2026/10/08 09:36

Prompt states the objective, Context supplies the background data, Harness governs the tooling and boundaries, Loop handles the iterative execution and validation, and Graph orchestrates task delegation, handoffs, and routing.

Read the full article：An AI Agent Is Basically an Agile Team: Demystifying 5 Core Concepts Through Everyday Software Engineering
Notes · 2026/10/05 15:50

Andrej Karpathy shared a tip:

Ask AI to explain things using ASD-STE100.

This is a writing standard developed by the aerospace industry for aircraft maintenance manuals:

One topic per sentence, one meaning per word—reducing misinterpretation and ambiguity.

He recommends applying only about 80% of it; applying it fully might actually leave out information.

The bottom line: AI shouldn't just get the answer right; it needs to be readable at a glance and easy to verify.

Read the full article：Making AI Answers More Readable: Karpathy on ASD-STE100
Notes · 2026/10/04 09:44

Decision Model / System One Model
- For text-only questions involving complex domain knowledge, use JEV
- For mixed text and images, ultra-fast responses, straightforward questions, and no domain expertise needed, go with Cloudflare Clef / Clef-flash
- For requirements combining the above, choose OpenAI Decisions API
Notes · 2026/09/26 10:57

Once the core architecture is in place

Have the agent take the self-contained, low-ambiguity tasks

and split them into multiple GitHub issues in the repo

Then build them and open PRs from ChatGPT connected to GitHub

Or even throw every AI tool you subscribe to at them — all guns blazing
Notes · 2026/09/26 10:57

Connect GitHub in ChatGPT on the web

Turn on the strongest model and max out reasoning effort

Have it create plan.md in the repo to hash out your most complex plan

After a few iterations, hand off to local Codex for long-running tasks

Double your quota, smoother flow, and a DIY boost to your DX
Notes · 2026/09/24 19:06

SaaS is being replaced by WaaS

Right now, building your own WaaS

could be your chance to ride the wave

Software as a Service

Workflow as a Service
Notes · 2026/09/24 18:54

From Prompt to Graph

AI agent workflows are closing the gap between intent and evidence

Upstream: intent engineering (e.g. jev)

Downstream: evidence engineering (e.g. better workflows / grok bots, which I'm building)