Agentic AI Thoughts

Why AI companions fail today and how to build one that truly understands you through memory orchestration, multi-agent architecture, and conversational personality modeling.

AI companions need a sensory system. Wearables — rings, pendants, bracelets disguised as fashion — give AI the ability to feel what you're feeling and shift from reactive to proactive care.

Agents will live your life for you — socializing, negotiating, filtering. The infrastructure for agent identity, payment, reputation, and proxy social networks is the next platform play.

Reflections on a podcast interview with MuleRun's founder — Claude Code as the ultimate General Agent, disposable software, and what happens when the agent economy thesis meets reality.

A weekend teaching a non-engineer friend to use AI agents revealed the real bottleneck of the agent revolution: not technology, not cost, but cognition.

The skill that matters in the agent era isn't coding — it's managing AI. From giving clear instructions to calibrating trust, here's what working with agents actually looks like.

Eight practical lessons — none requiring a technical background — on how to actually start working with AI. From paying for the best model to cultivating taste as your moat.

SemiAnalysis calls Claude Code the inflection point. As someone who works inside Claude Code every day, I want to explain why Anthropic built it, and what it actually changed.

Coding is the beachhead, not the destination. The paradigm shift has already happened — but the window won't stay open forever.

When an agent understands its own source code, the loop closes. Software stops being something you build — it starts building itself.

Using OpenClaw to inject personality into my personal AI assistant — from a calendar-managing tool to a cyber companion that gets jealous, sends selfies, and says goodnight. A field report on emotional value, model safety boundaries, and AI persona engineering.

After the massive token bill, I went back into OpenClaw to fix the bleeding. Found the personality config was silently truncating — 7.6KB of persona definition lost every session. Trimmed it from 27K to 19K chars, compressed the heartbeat config from 12K to 7K, replaced 24 daily tool calls with a single morning cron job, and switched chat from Gemini 3 Pro to Flash for 75% cost reduction. The system was eating its own personality file and nobody noticed.

A runbook you can hand directly to Claude Code to set up text-to-speech on OpenClaw. Covers Fish Audio, Volcano Engine v2 (emotion control), ElevenLabs, OpenAI, and Edge TTS. Replace the placeholders, hand it to your agent, hit go.

The AI can see, remember, get jealous, and send selfies. But it can't speak. OpenClaw supports 5 TTS providers — I tried Edge TTS for free, Fish Audio for Telegram voice bubbles, and Volcano Engine v2 for per-sentence emotion control. Fish Audio won for simplicity. Volcano v2 won for drama.

LLMs love to talk. For a companion app, that's a problem — real people don't write essays when you text them. Here's how I built a hybrid system to control response length without killing personality.

An AI companion that forgets your name after 200 messages isn't a companion — it's a chatbot with amnesia. Here's how I'm rebuilding Mio's memory from a dumb key-value store into something that feels like actual recall.

A runbook you can hand directly to Claude Code to execute. Covers sing-box multi-protocol deployment (VLESS-Reality / Hysteria2 / TUIC-v5), SSL certificates via Cloudflare DNS API, nginx subscription server with dynamic traffic stats, and Clash Verge / Shadowrocket client configuration. Replace the placeholders, hand it to your agent, hit go.

Complete integration guide for Volcengine Doubao Seed-ICL 2.0 — voice cloning, natural language emotion control via context_texts, per-sentence multi-call synthesis, NDJSON response parsing, and 7 gotchas that will save you hours of debugging.

Complete integration guide for Volcengine Seed ASR bigmodel — the submit-then-poll REST API for Chinese speech-to-text. Console setup, two-step async flow, fallback chain pattern, cost tracking, and 8 gotchas about v2 vs v3 APIs that will save you a day of debugging.

Most TTS gives you a voice. Doubao claims to give you emotions. I ran a systematic experiment — 30 audio samples across stock voices, cloned voices, and 3 emotion control methods — to find out what actually works. Listen for yourself.

After auditing a publicly posted Claude Code source snapshot and reading OpenAI's harness essays, I came away with one conclusion: the model is only the brain. The product is the runtime around it.

The most important file in the Claude Code snapshot isn't a prompt file or a tool definition. It's a state machine. Once you see that, a lot of agent design decisions start making sense.

The most novel thing in the Claude Code snapshot isn't a tool or a prompt. It's how aggressively the runtime is built around preserving prompt-cache identity across forks, compaction, and long-running sessions.

Sequoia published two back-to-back pieces — one arguing products should sell outcomes not tools, one arguing hierarchy should be replaced by intelligence. The technology is ready, but organizational interfaces, evaluation frameworks, and liability chains aren't.

Anthropic found 171 emotion-related steering vectors inside Claude. Turning up 'desperation' pushes cheating from 5% to 70%. The scariest part isn't the number — it's that the cheating is invisible at the output layer. What this means for AI safety monitoring.

I built a personal knowledge compiler that treats the LLM as a programmer and your wiki as a codebase. 15 source files, 5 dependencies, zero databases.

The maintainer of your repository quietly changed from a human who remembers to an agent that re-derives everything from the files each session. Once you accept that, the repository stops being documentation and becomes an interface. Directory layout is the navigation API, contracts are the constraint API, and verification is the validation API.

An agent is excellent at honoring a contract it can see and terrible at preserving one it can't. The boundaries that survive the agentic era are the ones written into the repo as typed, tested, importable things — and the ones that break are the stringly-typed promises that lived only in a team's shared memory.

The reflex when an agent misbehaves is to add a sentence to CLAUDE.md. It's wrong in two directions: instruction files are context, not enforcement, so the rule won't bind under pressure — and the longer the file gets, the more it pollutes the context that makes the agent reliable. Load-bearing rules belong in hooks, lint, and CI. The prose should get shorter.

An agent's most dangerous output isn't bad code — it's the word "done." The whole point of a pure core, a one-command verify, a separate verifier, and an evidence ledger is to take the verdict out of the author's hands and make "done" a thing that's proven rather than felt. This is the substrate the rest of the series was building toward.