Max

Obsessed with AI. Building AI-native software. New to open source.

[Max @maxaibuilds]

@hackernews Great news, welcome back!

[Max @maxaibuilds]

@OriSilver cool, will take a look! AGI

[Max @maxaibuilds]

Agent-friendly ≠ agent-native

Structured output, predictable errors, machine-readable help — table stakes.

The harder question: what abstraction should the agent operate through?

@karpathy @steipete — curious how you think about this.

[Max @maxaibuilds]

What should agent-native software interfaces look like?

Resource-oriented?
Intent-oriented?
Or both?
I wrote up my thinking — and would love to hear from people building agents that operate real software.
https://t.co/vYSLLjuT0E

[Max @maxaibuilds]

@LexnLin @AnthropicAI @snapsynapse @CogFoundryAI The full Aider redesign case study is here:

https://t.co/a8ni0itqZh

From the original site analysis to design exploration, implementation, validation, and the final result.

[Max @maxaibuilds]

What if an AI agent redesigned https://t.co/OGmCvansdw?
I built Redesign Lab, a @claudeai skill, to find out.
A real working redesign of Aider — an open-source, straightforward coding agent I really like.
#AIAgents #OpenSource
https://t.co/a8ni0itqZh

[Max @maxaibuilds]

@LexnLin @AnthropicAI @snapsynapse No. 5: loomloom from @CogFoundryAI.

It uses text-generate to write the case study — based only on verified facts from this run.

Actual cost: $0.0089 with #Gemini models.

https://t.co/FRIK1NXrRq

[Max @maxaibuilds]

@zhayujie Same starting point — your "the shape of the user changes: dev at a terminal, now an agent" is exactly the framing I've been writing from (CLIs for humans, APIs for programs, now agents). Learned a lot from it while redesigning my own CLI — thanks for open-sourcing this.

[Max @maxaibuilds]

@zbigniew I’ve been thinking about this one a lot. Your blog actually became a really important reference for my own work — I learned a lot from it and ended up using it as the checklist for designing my agent-native CLIs. So genuinely, thank you for putting it together.

[Max @maxaibuilds]

An agent thinks: run → inspect → retry → verify

But the CLI exposes:
work create
execution get
execution retry

Maybe agents need intent first, resources for state and recovery.

Which abstraction should agents use?

[Max @maxaibuilds]

I've been thinking about a shift in software interfaces.
CLIs were designed for humans.
APIs were designed for programs.
Now we're designing interfaces for agents.
I suspect “make the output JSON” is only the beginning.
What happens when agents become the primary operators?

[Max @maxaibuilds]

@undefinedKi @gippp69 Good question. Coding makes the gap obvious because results are binary.
Outside of code it may be quieter, but generation and verification still pull on different strengths. Would be interesting to test the same pairing on research or open-ended tasks.

[Max @maxaibuilds]

@antpalkin The sixth step is the real gap.
Everyone is getting better at generating and testing strategies. Almost no one is building the part where the system actually remembers what already failed and stops repeating it.
Been thinking about building something in that direction.

[Max @maxaibuilds]

@chesny Clean shift from linear creative work to parallel multi-agent production.
Keeping the human only at the final gate while 6 agents run in parallel is interesting — that hits real limits as volume grows.
Curious how @Creatify_AI structures those workflows under the hood.

[Himanshu Kumar @codewithimanshu]

A trader built a QUANT bot using Claude Fable 5.

I saw it & asked Claude Fable 5 to do something wild:

Analyze undervalued markets on Polymarket to find arbitrage opportunities and detect the wallets exploiting them so I can copy them.

Turned $2K -> $11K overnight.

It monitored over 1,000 wallets.

I realized something fast:

There are arbitrage bots I can't compete with without knowing how to program.

But I can find them.
And copy them.

Fable 5 built me a monitoring terminal connected to a copytrading bot on Telegram.

Not a script.
Not even a bot.

An AI agent that improves every time it finds a new wallet.

It analyzes the behavior of each wallet:

→ How it operates
→ Its arbitrage patterns
→ The size of its positions
→ The exact moment it enters

70% hit rate right now.
7 wallets being copied out of 500+ monitored.

The bot never stops.
Never bets.
It just follows the math and seeks profit.

Most people are trying to predict the market.

This system just identifies who's already winning and copies their exact moves.

No guessing.
No emotions.
No trading experience needed.

The entire edge is pattern recognition + execution speed.

While traders debate which way BTC will move, this agent is already positioned in the same trades as wallets pulling $10K+ per day.

You only need:

→ Claude
→ A device
→ 1 hour per day

The system runs autonomous after setup:

→ Monitors 1,000+ wallets continuously
→ Detects arbitrage patterns in real-time
→ Copies winning positions automatically
→ Improves selection logic with each trade

No manual trading.
No chart reading.
Just identifying proven edges and following them.

💡 I'm sharing the complete monitoring terminal + copytrading bot setup for free.

24 hours only.

To get it:

1️⃣ Comment "Fable"
2️⃣ Like and Repost
3️⃣ Follow @codewithimanshu ( So i can DM you)

I'll DM you everything.

[Max @maxaibuilds]

@codewithimanshu Fable