equip a coding agent

You want to learn what AI is good at before trusting it with anything. At this level the model only sees what you paste into it, which is exactly why it is safe.

No wiring at all. You carry context to the model by hand and carry its answers back by hand. You are the hands, the eyes and the memory.

Shapes on the map: Chat-only (web chatbox) · IDE copilot (inline AI)

Any capable chat model works. A local model is worth installing even if it is weaker: it teaches you what the model does versus what the wiring does.

• VS CodeEditor with inline AI
• OllamaFree local model to experiment with

• Never paste keys, passwords or personal data into a chat window.
• Read every line the model writes before you run it; you are the only test suite.

You catch yourself copy-pasting files into the chat. The work is now bigger than a single answer: build this, fix that, deploy it, prove it works.

The agent gets hands inside one repository: it can read files, run commands and see the output. Git sits underneath as the safety net, and a skill.md file carries memory from one session to the next.

Shapes on the map: Terminal agent (this site's shape)

Pick a terminal agent that shows you every command before it runs and can be limited to a folder. Pay-as-you-go API is fine until you use it daily.

• Claude CodeTerminal agent — reads, runs, fixes
• GitUndo button before the agent touches files
• GitHubRemote backup and CI checks
• ObsidianWhere skill.md memory lives

• Commit before you let the agent act, so any change is one command away from undone.
• Ask for evidence, not a claim: test output, logs or a fetched page — not the word done.
• Keep secrets in environment variables the agent can use but never prints.

The agent now does real work that touches other systems — a live website, a database, a browser — and the work keeps going when you are not at the keyboard.

The terminal agent gains plug-in tools through a standard protocol, and a chat channel lets you steer it from your phone. Every tool is a separate permission, and anything irreversible still waits for your yes.

Shapes on the map: Terminal agent (this site's shape) · Terminal agent + phone channel

Prefer tools that speak the open protocol over custom glue, so they keep working when you switch agents. Give each one the least access that does the job.

• Model Context ProtocolPlugs in browsers, databases, services
• Telegram Bot APIPhone channel to drive it from anywhere
• WhisperVoice notes become instructions
• GitHub CLIThe agent's door into GitHub

• Grant each tool least privilege: read-only unless writing is the point.
• Anything irreversible — deleting, paying, publishing, sending — requires your explicit confirmation.
• Log every action the agent takes through a tool, so you can reconstruct what happened.
• Set a hard monthly budget on the model account.

One context can no longer hold the work — a migration, an audit, parallel research — and several long jobs are queuing behind a single session. Only move here once one agent is genuinely saturated.

An orchestrator splits the job and hands pieces to worker agents, each in its own isolated copy of the code. They hand work back through files and reviewed changes, and a separate agent checks the results before anything merges.

Shapes on the map: Multi-agent fleet (orchestration)

The hard part is not spawning agents but merging their output. Choose tooling that gives each agent its own workspace and makes every result reviewable.

• Claude CodeOrchestrator and workers
• GitHubEvery agent's work lands as a reviewed change
• NotionShared task board agents read and write

• Give every worker its own working copy; two agents editing one folder corrupt each other's work.
• Have a different agent verify the result than the one that produced it.
• Set hard limits on time and tokens per agent; a stuck agent should stop, not loop.
• A human approves every merge into the main line.