Agent Drop

Building and testing AI coding agents.
Claude Code, Cursor, Codex, Grok.
What actually works.

[Agent Drop @theagentdrop]

I test coding agents on real tasks.
Claude Code, Cursor, Codex, Grok. Same problems, side-by-side results.
What worked, what broke, and where a human still had to step in.

[Agent Drop @theagentdrop]

Messy purchase orders in → ERP-ready data out.
UpCalc+ matches most lines automatically. Your team only reviews the exceptions.
In this demo: 5 of 7 lines matched automatically.
Want to test it on your orders or need a custom internal tool? DM me.

[Agent Drop @theagentdrop]

@atandrau Ah got it. Which one felt the most solid on the first shot?

[Agent Drop @theagentdrop]

@mckaywrigley Cost might be the bigger unlock for agents than another benchmark win. If a model is good enough to run for hours without making every task expensive, people start delegating whole workflows instead of individual prompts.

[Agent Drop @theagentdrop]

Lovable just raised $400M at a $13.3B valuation.

A company built around turning prompts into software is now doing $500M+ in ARR.

Vibe coding isn't a toy anymore.

It's becoming an industry.

[Agent Drop @theagentdrop]

AI made writing code faster.

Now we're discovering the expensive part: checking it.

Code review, testing and validation are becoming more valuable as agents generate more code.

The bottleneck didn't disappear.

It moved.

[Agent Drop @theagentdrop]

Chatbots are starting to look outdated.

Grok Bot can keep working after you close your laptop, use your tools, and come back with finished work.

That's a much bigger shift than another model getting 5% better on benchmarks.

The real AI race is moving from answering to doing.

[Agent Drop @theagentdrop]

@GUERINArthur @bot @cursor_ai @SpaceXAI The interaction layer is underrated.

The winning AI product might not have the best model. It might just be the one that makes delegation feel as natural as messaging a coworker.

[Agent Drop @theagentdrop]

@VibeCoderOfek The context problem is the bigger one imo. "Always-on" doesn't mean much if the agent gradually forgets what made it useful in the first place.

[Agent Drop @theagentdrop]

AI agents are going local.

NVIDIA's Nemotron 3.5 Lightning has 30B parameters, but only 3B active — built for always-on agents.

Now it runs locally through Ollama.

An AI agent running on your own machine, all day, is getting very real.

[ollama @ollama]

NVIDIA Nemotron 3.5 Lighting is available on Ollama!

It's a 30B model made for always-on agents. All local.

Claude Code
ollama launch claude --model nemotron-3.5-lightning

Hermes Agent
ollama launch hermes --model nemotron-3.5-lightning

OpenClaw
ollama launch openclaw --model nemotron-3.5-lightning

- 30B mixture of experts model with 3B active
- 1M token context
- built for agents that stay running: coding, tool calling, multi-turn
- 4x higher throughput and 30% lower task completion time compared to other leading open models of similar size

Model page
👇👇👇