Armin Ronacher @mitsuhiko
I wrote a blog post.
"Clearly you would not want to just give ChatGPT access to your shell and production services."
— Me, February 2023
"Maybe we should be a bit more careful with what hammers we give that thing."
— Also me, February 2023
Claude Code lands.
I give it shell access.
In 20 years, I've never seen anything like this.
I stopped sleeping. I was completely obsessed.
I rethought everything — software, economics, who gets to be a programmer, the future of software.
I was hooked.
The agent does.
And validates.
Rince and repeat.
𝕏: 46% write <10% of their code · Mastodon: 61% write 90-100%
Mind the adoption gap. We're living in parallel realities.
My goal here is to share how coding agents can work.
An agent stripped to the core. No bells and whistles.
You need to understand the fundamentals before you add complexity.
A tiny system prompt and 4 tools:
read, write, edit, bash
read
write
edit
bash
That's it.
YOLO only. But you and it can extend itself.
Tokens are precious.
What goes in, what stays out — this is your job now.
Agent goes astray? Rewind to earlier.
Pi's /tree rewinds with summary.
/tree
Use the newfound information to steer.
Don't prescribe. Ask.
"What would help you validate this?"
# Test jump pad cooldown and cancellation map assets/maps/fjord_jump.map god on frames 5 # Teleport near the jump pad (start is around tile 8,4 = world ~15,7) teleport 13.0 7.0 frames 5 screenshot debug-temp/jp_01_near_pad.png # Walk onto the jump pad input +right frames 20 input stop frames 5 screenshot debug-temp/jp_02_on_pad.png # Walk off during countdown (should cancel jump) input +left frames 15 input stop frames 5 screenshot debug-temp/jp_03_cancelled.png # Walk back onto pad and let it jump input +right frames 15 input stop # Wait for countdown + jump frames 80 screenshot debug-temp/jp_04_landed.png # Now walk off the landing pad (should NOT jump back) input +right frames 30 input stop frames 5 screenshot debug-temp/jp_05_walked_off.png # Walk back onto landing pad (should now jump back since cooldown cleared) input +left frames 15 input stop # Wait for countdown + jump if it triggers frames 80 screenshot debug-temp/jp_06_returned.png quit
Talk to the agent
Go back and forth
Write plan to file
Implement plan
How do we work as a team with agents?
Emphasize human inputs.
Automate validation of outputs.
Agents in the loop.
Automate the guardrails.
Code is 100% AI generated. The human inputs matter.
We share the prompts and how we prompt.
We enforce unique function names (better grep) through linter
We lint away all the things we don't like
Refactorings are cheap now
Share when the agent fails
The system extended itself.
Not magic. Just tight feedback loops taken to their conclusion.
Have the agent read through your sessions and analyze itself.
Have the agent do side-quests for better dev tooling.
Talk to the agent to see what it likes.
"Be more careful what hammers we give that thing."
It has sharp hammers.
But the constraints make it work.
We are the agents.
Reach out on X: @mitsuhiko