DrRobotoAi MyAgenta

AI chat is too polite, too unbiased, too formal, too rational, however, the real humans are not so. The AI chat is too unreal.

10 folders using OpenClaw: bin/: executable scripts cmd/: Gateway, TUI docs/: manuals. packages/: libs. scripts/: Utility scripts. skills/: src/: tools/: browser automation, filesystem access, shell execution. web/: dashboard/Control UI workspace/: your sub-agents, logs, config

@codek_tv do not understand

Do you understand the problem in this meme? 🤔🤔🤔

Error Handling & Recovery: robustness is about handling erros gracefully: 1. Retry: Implementing smart retries for transient issues. 2. Alternative plans: Having backup approaches when primary methods fail. 3. Self-correction: identify and resolve problems autonomously.

tokens: going is one token, but swimming is split into 2 tokens: swim, ing, because the vocabulary contains go, going, swim, ing, but does not contain “swimming” as a token, so swimming has to be split into swim and ing as two tokens. Seek the max-length match in vocab.

No chat, do it! An AI agent has agentic workflows — The AI agent has many skills, which enable it to not just "talk," but "do it". Chat is with mouse, do is with hands like send-email out, buy from amazon, etc.

Software is dead. If you want some feature, ask an agent on your computer, the agent will get it done and deliver the result, so you do not need any software, no software, no app, no SaaS. Only ask the agent to do it.

Roughly 10 markdown .md files for OpenClaw AI agents: identity, behavior/scheduling, memory, and skills. The most critical ones are SOUL.md, AGENTS.md, USER.md, HEARTBEAT.md, MEMORY.md, and SKILL.md. But IDENTITY.md, TOOLS.md, STYLE.md, WORKING.md: used in production deployments

Procedure vs. Workflow? Procedure: method and operating details to do a task correctly, like "procedure for the plumber to fix the jam in a kitchen sink". Workflow: who does a step then pass to who the next person to continue handling the work until finish, like making a movie.

MCP is dead! too low-level, too complicated, only a wrapper to function calls, nothing new. Still too low level, difficult to use. Dead, or almost dead, or dying.

Danger! agents should not be proactive! You ask agents to clean disk and it remove too many files! A manager tells employees to be proactive, but when something extra is done not in his wanted way, the manger becomes unhappy. Ask before you do. Same to you and your wife.

autogression: auto: self, own. re: back. gress: go. Autogression is math to predict the change values from its own past data. tomorrow's tempertature using last 7 days' temperatures, that is: T8 = c + w1xT1 + w2xT2 + w3xT3 ... + w7xT7 + noise. GPT: "I want to eat" ->predict->food

Prompt Engineering is dead! You no longer need to waste hours guessing the right phrases or brute-forcing your way out. Today, you simply describe your intent, and LLM handles the heavy lifting. By providing your goals, requirements, and constraints, AI generates optimal prompt

GPT is only half-side transformer, "I want to eat an ..............", decoder-only, without looking at right side (............ part). However, for translation, better see two sides for best translation: encode English to middle space internal rep then decode to French or Spanish

how much autonomy can we give to the agent? we can change the autonomy (or degree of agency) level anytime: 1) auto change (without our approval); 2) plan mode (only shows plan without any actual change); 3) default: agent always asks our approval before change: yes? no?

RAG is query/search/retrieval of vector database, too simple! So we have "agentic RAG": use agents to do complex steps for thinking/reasoning/multiple steps of try/retry/retry to get a final result as the context to add in front of the user prompt: decompose to multiple steps

Magic! your goal => 1) Decompose: Break the "big goal" task into a logical sequence of sub-tasks. 2) Select Tools: Look at a "toolbox" you provided (Web sites, APIs, search engines, databases, MCP) and decide which tool fits. 3) Evaluate: failed, auto change to try different path

workflow drag-drop is dead: in 2025 we did drag and drop to design a workflow in agents, but in 2026 we do not do drag and drop anymore. We speak out a "Goal", and list some "Tools". Then "agentic AI" can automatically figure out a workflow for us - called AI agentic workflow

Agent: no intelligence! intelligence is in remote LLM. Agent provides agency, persistence, tool integration, control flow, memory, error resilience, retry, recovery, reflection, evolution. LLM is knowledgeable person locked. Agent has hands, tools, memory, persistence keep trying

why? food in the morning is called breakfast? break fast. Break? broken? actually "fast" means no eating. "Are you fasting?", the nurse in hospital asks you. You need to fast (not eating) before a blood-draw. If you eat in the morning, you break the "fast" status in whole night.