I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
I gave GPT-6.2 Astra and Fable 5.2 the same prompt to make a 10-second cowboy toy film in Blender.
GPT-6.2 Astra → ~2 hours, ~60K generated tokens,
Fable 5.2 → ~20 hours, ~150K–300K generated tokens
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
This is f*cking insane. Fable 5.2 vs GPT-6 Astra:
Same prompt. Same logging truck challenge.
Fable 5.2: $6 and 2 minutes.
Astra: $96 and 26 minutes.
Fable 5.2 → Opus 5.5 → Blender MCP → GPT-6 Astra
Fable built a Hollywood-level mountain scene with massive trucks, explosive dust, cinematic lighting and incredible environmental details. Astra delivered a rejected PS2 game asset 😭
In this comparison, Fable was 16x cheaper and 13x faster than Astra.
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
Gemini 4 Hydrogen vs Gemini 4 Carbon vs GPT-6.1 Sol vs Muse Spark 1.3 🧊
Google vs OpenAI vs Meta. One prompt. 🤯
Same prompt: Cathedral of Ice
Gemini 4 Hydrogen: $4.27 · 1h 23m 40s
Gemini 4 Carbon: $8.83 · 2h 39m 10s
GPT-6.1 Sol: $3.11 · 54m 19s
Muse Spark 1.3: $5.56 · 2h
GPT-6.1 Sol was the cheapest and the fastest of all four. 😱
Almost 3x cheaper and almost 3x faster than Gemini 4 Carbon.
Which one would you pick? 👀
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
Haters said "nobody builds in Three.js"
Okay... explain this:
how can a coded model be on par with a 3D mesh cooked in Blender??
Top is Three.js
Bottom is Blender
water, sky, and all the simulator mechanics = built via Opus 5.5 + Fable 5.2 too
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
Fable 5.2 is absolutely f*cking insane for film animation.
i made this 2-minutes animated short with Fable 5.2 without writing a single line of code..
and it's better than 90% of video models.
> pure three.js, zero video model
> 8 worlds, 16 shots, 3,000 code-rendered frames
> music + sound effects synthesized in code
> built by a team of 17 parallel AI agents
Find full prompt below
then, read the full guide on motion design with Fable 5.2
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
🚨Your agent isn't dumb. Your tools are.
Anthropic's engineers wrote down how they build the tools Claude uses. The lesson: most agents break at the tool layer, long before the model.
So I applied their 5 rules to a wallet agent:
→ 4 API calls it kept chaining wrong, rewritten as 1 tool
→ token symbols instead of raw 0x addresses
→ concise replies by default: 72 tokens instead of 206
→ errors that tell the agent how to fix the call
→ one prompt that lets Claude rewrite its own tools from failed runs
Your agent never reads your code. It reads tool names, descriptions and outputs. That text is part of your harness.
The rest of the harness is in my article ↓
@OxEnsooo This is something people often overlook. Everyone talks about better models, but the tools around them matter just as much. Small changes like these can make a huge difference. Great work on this!
@AI_cortexx Kinda scary when you think about it. We keep building smarter AI tools, but they still depend on services that can go down at any moment. One outage and suddenly half your workflow is gone.
@lez99999 This is getting crazy 😂 A whole team of agents working together while you just keep an eye on things. Definitely curious to see where this goes.
I just solved "drone engineering" education forever with Fable 5.2...
no need to break a $2,000 drone to see what's inside, just open my drone lab:
• 3 real drone types in 3D: FPV racer, farm sprayer, camera drone
• exploded view: every part flies out with what it does
• 24 parts: carbon frame, 4-in-1 ESC, spray pump, RTK GPS, dual-lens gimbal...
• then hit "fly": speed run, spray run over a corn field, flight test
every 3D model is generated in code, no ready-made 3D files or photos
article for your promt below:
I've been thinking about why AI agents fail at tasks they should easily handle.
Maybe we're too focused on making models smarter and not enough on building better systems around them.
Here's what I found. 👇 https://t.co/jHTfBFUP4T
@lez99999 This is the part I love about the AI space. People are sharing actual workflows instead of just talking about what AI might do. Definitely going to check out some of these repos.