Introduction to Claude Code and Codex

From autocomplete to agents

AI-assisted coding and autonomous driving have followed parallel tracks. Coding assistance moved from IDE autocomplete in the 1990s through IntelliSense, TabNine, and GitHub Copilot to terminal agents like Claude Code in 2025; driving moved from cruise control through lane assist and Tesla Autopilot toward full autonomy. Both tracks tell the same story: assistance that once reacted keystroke by keystroke now plans and acts over long horizons.

The autocomplete era looked like this: you type a prefix, the IDE ranks completions from documentation, and you pick one.

# User types:  "model = nn.Tra" + Ctrl+Space
model = nn.Tra|

# IDE intelligent suggestions
Transformer(d_model, nhead, ...)   ← selected
TransformerEncoder(encoder_layer, num_layers)
TransformerDecoder(decoder_layer, num_layers)

# Auto-completion with parameter hints:
model = nn.Transformer(
                d_model=|, # int:  model dimension
                nhead=8,   # attention heads
                )

What changed with agents is the position of the human in the loop. In the traditional arrangement, the human sits between the AI and the environment: the model answers a query, the human performs the action and relays the result back.

Agents collapse that relay. The model acts on the environment directly, reads the feedback itself, and reports results; the human supplies intent and oversight rather than keystrokes.

Claude Code: an agent in the terminal

Claude Code is that agent loop packaged as a terminal program. You talk to it in natural language, as simple as you are a five-year-old child; it reads files, edits them, runs commands, and reports back.

✱ Welcome to Claude Code!

  /help for help, /status for your current setup

  cwd: /Users/reself

> Hello World!

Memory

The agent keeps three layers of memory, and knowing where each lives saves repeated explanations:

• Global memory: .claude/CLAUDE.md in your home directory;
• Local memory: CLAUDE.md within each project root;
• Session context: maintained during active conversations.

Permissions

Every tool call passes through a permission layer you configure once and refine over time. Allow the harmless, ask on the consequential, deny the dangerous:

"permissions":  {
    "allow":  ["Bash(ls:*)", "Bash(git status:*)", "python:*"],
    "ask":    ["Bash(git commit:*)", "Bash(sbatch:*)"],
    "deny":   [],
    "additionalDirectories":  ["/mnt/..."]
}

Custom commands

Frequent tasks become slash commands: write a markdown file describing the task, store it in .claude/commands, and it becomes invocable. Both global and local commands are supported. A cluster-status checker, for example:

# Check SLURM Jobs            → saved as check-slurm.md, invoked as /check-slurm
Check my current SLURM job status and queue information.
Direct squeue command without bashrc dependencies.
## Command
squeue -u <your-id>

Statusline

The statusline turns the terminal footer into a dashboard. Mine shows running costs, the current todo, live SLURM jobs, and recently finished runs:

CLAUDE COSTS
Today    $8.56
Total    $231.53

TODO LIST
increasing batch size?

SLURM JOBS
loopb4p4s416e4               8:17:54    H100x4
loopb4p4s416e4_le_iiadd_ni   QUEUE #4   H100x4

RECENT COMPLETED
loopb4p5s216e5               8h ago

Image input

Claude Code accepts images: paste a screenshot with ctrl+v and reference it in the prompt.

> [Image #1] Please analyze this screenshot
and help me recreate the layout in CSS.

MCP: Model Context Protocol

MCP is the universal standard for connecting AI systems with external tools and data sources, with pre-built servers for Google Drive, Slack, GitHub, Git, Postgres, Puppeteer, and more.

Agents

Recurring specialized work deserves its own agent. The /agents interface asks what the agent should do and when it should be used; Claude Code then invokes it automatically when a task matches.

> /agents
Create new agent
Describe what this agent should do and when it should be used (be
comprehensive for best results)

the agent will help me polish the papers by looking a latex-based repo

Press Enter to submit  ·  Esc to go back

Codex

Why run a second agent at all, and how does Codex differ from Claude Code? The short answer from daily use: they have different temperaments, and the difference is exploitable.

Setup

Launch Codex in the terminal with the codex command, set /model to gpt-5 high for best performance, and configure /approvals to read-only for safety. The three approval modes:

1. Read Only: Codex can read files and answer questions; it requires approval to make edits, run commands, or access the network.
2. Auto: Codex can read files, make edits, and run commands in the workspace; it requires approval to work outside the workspace or access the network.
3. Full Access: Codex can read files, make edits, and run commands with network access, without approval. Exercise caution.

Sharing memory across both tools

Outlook

Three reasons to be optimistic about Codex specifically:

• Reasoning: reasoning capabilities with test-time scaling by RL;
• General agent: OpenAI has already built a general agent in web. Web is the next new terminal;
• Multi-modal: OpenAI cares more about multi-modal input and output.

Future agents will cover the vast majority of multi-modal contexts in our daily lives as prefill, and execute through extremely complex environmental interactions. Therefore, we will simultaneously have ultra-long context capabilities. How to handle conflicts between contextual information and pretrained knowledge will also become a key focus for future model development.

Fun fact: the slide deck behind this post was collaboratively created, refined, and polished using Claude Code and Codex.