Lexicon

Large Language Model: a model trained on large text corpora. It predicts the next token with high skill. It answers, drafts, translates, and codes — it does not “understand” the way a person does.

Strong next-token prediction is not agency by itself.

LLM plus tools plus memory. A chat product answers. An agent acts: files, terminal, browser, messages, scheduled work.

Brain + hands + memory = agent.

The unit of model I/O and, for cloud APIs, cost. Roughly three-quarters of a word. Local inference shifts cost to hardware and time.

Cloud bills tokens; local bills hardware and patience.

An API key authenticates a program to a provider. A .env file holds secrets outside the repository. A leaked key is unauthorized spend in your name.

Keys in .env. Never commit .env.

Runtime is where the process runs: PC, VPS, container. Harness is the frame that turns a model into an agent: loop, tools, context, memory, approvals, gateways.

Runtime = where. Harness = how it becomes yours.

A skill is a reusable procedure (often SKILL.md). A plugin adds product capability. MCP is an open standard for connecting external services through one interface.

Skills describe how; MCP describes what you plug in.

Local models run on your machine — data stays, you pay hardware. Cloud models run at a provider — pay per use, often higher capability, data leaves. A good harness lets you change this choice without rewriting the agent.

What matters is an explicit choice, not ideological purity.

A scheduled job. Example: each weekday at 08:00, fetch headlines, summarize three items, deliver to Telegram. Work that continues without you in the chair.

Harness + cron = recurring work without a live session.

Everything the model can see in the current session. Large and volatile: close or prune the session and it fades. Treat it as RAM, not permanent storage.

Context is session RAM; durable memory is what you write to disk.

Models must emit tokens. When evidence is missing, they complete the pattern with plausible text. Grounding in files and sources reduces, but does not eliminate, that risk.

Unknown → invention. Prefer files you can open.