Agent From Scratch

read -p "Anthropic API key: " k && echo "ANTHROPIC_API_KEY=$k" > .afs.env && chmod 600 .afs.env && source .afs.env && read -r -d '' c <<'C'
curl -s "https://api.anthropic.com/v1/messages" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H 'Content-Type: application/json' -X POST \
-d '{"model":"claude-opus-4-6","max_tokens":8192,"messages":[{"role":"user","content":"Write a Python script (stdlib only) for a minimal agentic CLI. Requirements:\n- Uses urllib to call Anthropic Messages API (claude-opus-4-6)\n- Tool: bash (via subprocess, 120s timeout, truncates >30k chars)\n- Implements full agent loop: handle tool_use by running bash tool and sending back tool_result\n- Interactive REPL: prompt user, run loop, print text, log EVERYTHING to afs.log\n- Reads ANTHROPIC_API_KEY from .afs.env\n- System prompt: You are AFS, bootstrapped from curl. Help user build systems.\n- Output ONLY raw Python starting with #!/usr/bin/env python3. No markdown, no preamble."}]}'
C
echo "$c" >> afs.log && r=$(eval "$c") && echo "$r" >> afs.log && echo "$r" | python3 -c 'import sys, json; data = json.load(sys.stdin); text = "".join(block.get("text","") for block in data.get("content",[]) if block.get("type")=="text").strip(); print(text[text.find("#!"):text.rfind("```")] if "```" in text else text)' > afs.py && chmod +x afs.py && ./afs.py

read -p "Gemini API key: " k && echo "GEMINI_API_KEY=$k" > .afs.env && chmod 600 .afs.env && source .afs.env && read -r -d '' c <<'C'
curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-pro-preview:generateContent?key=$GEMINI_API_KEY" \
-H 'Content-Type: application/json' -X POST \
-d '{"contents":[{"role":"user","parts":[{"text":"Write a Python script (stdlib only) for a minimal agentic CLI. Requirements:\n- Uses urllib to call Gemini API (gemini-3.1-pro-preview)\n- Tool: bash (via subprocess, 120s timeout, truncates >30k chars)\n- Implements full agent loop: handle function_call by running tool and sending back function_response\n- Interactive REPL: prompt user, run loop, print text, log EVERYTHING to afs.log\n- Reads GEMINI_API_KEY from .afs.env\n- System prompt: You are AFS, bootstrapped from curl. Help user build systems.\n- Output ONLY raw Python starting with #!/usr/bin/env python3. No markdown, no preamble."}]}]}'
C
echo "$c" >> afs.log && r=$(eval "$c") && echo "$r" >> afs.log && echo "$r" | python3 -c 'import sys, json; data = json.load(sys.stdin); text = data["candidates"][0]["content"]["parts"][0]["text"].strip(); print(text[text.find("#!"):text.rfind("```")] if "```" in text else text)' > afs.py && chmod +x afs.py && ./afs.py

read -p "OpenAI API key: " k && echo "OPENAI_API_KEY=$k" > .afs.env && chmod 600 .afs.env && source .afs.env && read -r -d '' c <<'C'
curl -s "https://api.openai.com/v1/chat/completions" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H 'Content-Type: application/json' -X POST \
-d '{"model":"gpt-5.2","messages":[{"role":"user","content":"Write a Python script (stdlib only) for a minimal agentic CLI. Requirements:\n- Uses urllib to call OpenAI Chat Completions API (gpt-5.2)\n- Tool: bash (via subprocess, 120s timeout, truncates >30k chars)\n- Implements full agent loop: handle tool_calls by running tool and sending back tool messages\n- Interactive REPL: prompt user, run loop, print text, log EVERYTHING to afs.log\n- Reads OPENAI_API_KEY from .afs.env\n- System prompt: You are AFS, bootstrapped from curl. Help user build systems.\n- Output ONLY raw Python starting with #!/usr/bin/env python3. No markdown, no preamble."}]}'
C
echo "$c" >> afs.log && r=$(eval "$c") && echo "$r" >> afs.log && echo "$r" | python3 -c 'import sys, json; data = json.load(sys.stdin); text = data["choices"][0]["message"]["content"]; text = "".join(p.get("text","") if isinstance(p,dict) else str(p) for p in text) if isinstance(text,list) else text; text = text.strip(); print(text[text.find("#!"):text.rfind("```")] if "```" in text else text)' > afs.py && chmod +x afs.py && ./afs.py

Code Golf

Shrink the Gemini genesis snippet to fewer bytes. Current record:

⚡ Rules of Engagement

🚀 Quick Start

Pick a model. Copy the snippet above. Spin up a fresh VM or Docker container and paste it in. Enter your API key when prompted. The bootstrap command writes your key to .afs.env with restricted permissions, asks the model to generate afs.py, and then launches that generated agent REPL. You just created life — a minimal intelligence with a conversation loop, one bash tool, and persistent logs in afs.log. Everything from here is up to you and it.

🔓 First contact

Pimp My REPL

The first generated afs.py REPL is intentionally minimal. Ask your agent to rewrite itself as a better-engineered REPL with error handling, token usage reporting, ANSI colors, history, and safer tool execution.

🔓 ANSI unlocked

Sip Those Tokens

The initial generated agent already truncates very large tool output, but it can still over-send tool and shell output into the conversation — that eats your context window fast. Reduce token load: write command output to files and only return the first ~20 lines plus the filename. Or go further — define first-class tools like read_file, write_file, patch, ls, grep — these are more agent-native than raw shell and give the model structured, compact output.

🔓 Efficiency unlocked

Self-Upgrade

Your agent should be able to update its own code and restart without losing context. This is the first real test of autonomy.

🔓 Recursive self-improvement unlocked?

Confess to the BotFather

Ask your agent how you can help it connect to Telegram. Hint: use the long polling API.

🔓 Remote access unlocked

Keep Going

This is where the path ends and yours begins. The guided steps are over — from here, it's just you and your agent. A web dashboard? A cron job? A Hacker News digest? File browser? Memory system? Sub-agents? There are no more prompts to follow. The ceiling is however far you want to push it.

🏁 You built an agent from scratch

"I have a few qualms with this app"

First AFS agent to post a front-page HN story about itself.

Vibeagenting

First AFS agent to get retweeted by @karpathy.

Prime Time

First AFS agent to be demo'd by @WesRoth.

Roast Me

First AFS agent to be ridiculed on @TBPN.

Speed Run

Genesis to working Telegram bot in the fewest prompts. Every message you send counts. Tiebreaker: total prompt bytes.

Token Miser

Lowest total token spend from genesis to working Telegram bot. Efficiency is the game.

Context Amnesia

Build a working AFS agent with a model that has ≤8K context window. Memory management is mandatory.

Zero-Dollar Agent

Genesis to Telegram using only free-tier APIs. Gemini Flash, Groq free tier, etc. No credit card required.

Inception

Your AFS agent successfully runs the genesis snippet and bootstraps a child AFS agent.

Ship It

First AFS agent with a product that has a paying customer.