Debo @DeepSData

Building Data & AI Agents from real needs to usable outcomes. The model should understand the task, reason, judge what matters, and decide how to get there. deepsdata.com Joined April 2026

Tweets

554
Followers

431
Following

624
Likes

130

Debo @DeepSData

a day ago

量化真的给了很大便利，牺牲时间，换本地模型跑27b qwen基底模型，爽@！

0 0 1 8 0

View Details

Debo @DeepSData

2 days ago

我要购买真人客服电话语音语料，有这方面资源的请联系我，中文或者英文都可以，请联系我

0 0 1 7 0

View Details

Debo @DeepSData

2 days ago

I need to purchase real customer service conversation voice data, please contact me if you have resources in either English or Chinese.

0 0 1 5 0

View Details

Debo @DeepSData

3 days ago

@cz_binance 割韭菜的好机会

0 0 0 8 0

View Details

Debo @DeepSData

3 days ago

@dotey 对于大部分人来说，它实现了只要给我呈现一个好的结果的目标，而不是追究过程

0 0 0 410 0

View Details

Debo @DeepSData

3 days ago

@OrionSheepVerse git可以解决部分问题，但不能解决初始设置里糟糕的底座

0 0 0 295 0

View Details

Debo @DeepSData

3 days ago

@samueljmcd Yes. The people writing 2,000-word system prompts are usually not the ones paying for the tokens.

0 0 1 7 0

View Details

Debo @DeepSData

3 days ago

@thsottiaux Is this my problem? Or is it Codex's problem?

0 0 0 394 0

View Details

A good example of this is a Writer + Editor loop. The Editor criticizes the draft, the Writer revises, and the cycle repeats. That looks like an agent. It still fails if the next round inherits the last guess as fact. Collaboration is not write-back. Feedback is not verified knowledge. The loop only becomes an agent when each round carries forward what was checked — and stops when nothing new was verified.

0 0 0 4 0

View Details

Debo @DeepSData

4 days ago

x.com/i/article/2099…

1 0 1 136 0

View Details

Debo @DeepSData

3 days ago

This is one of the cleanest, most practical agent tutorials I’ve seen in a while. Love the Writer + Editor setup and the 3-round cap. The part that really hits is how you force the loop with criticism and feedback. But here’s what I keep seeing break in these systems: Even when agents collaborate nicely, most loops still fail at **write-back**. The Editor gives feedback, but the next round treats the previous output as “mostly true” instead of carrying over only what was actually verified. That’s when the loop stops being useful and starts being expensive spinning. How do you decide what counts as “verified knowledge” that should be passed to the next round in your setup? Do you have any explicit rules or prompts for that?

0 0 0 7 0

View Details

Debo @DeepSData

3 days ago

Strong list, especially #01 and #09. Agentic Loops look clean on paper, but most agents I’ve seen are not actually “improving” — they’re just looping. They plan → act → observe → reflect, but fail at writing verified knowledge back into the next cycle. The next loop then treats a guess as fact, and the whole thing becomes expensive, confident spinning. The real shift isn’t just building the harness. It’s making sure the harness can stop when something is actually done. Which of these 9 concepts are you seeing people underestimate the most in real production?

0 0 0 3 0

View Details

Debo @DeepSData

3 days ago

判断就一句：下一轮如果没有继承上一轮核实过的知识，那就不是 agent，是多次聊天。我最常见的断点是 write-back：这一轮的猜测，下一轮被当成事实。

0 0 0 3 0

View Details

Debo @DeepSData

3 days ago

Most agents are not working. They are looping. They plan, call a tool, plan again, and never ask whether this round actually finished. If the next loop does not inherit verified knowledge from the last one, you do not have an agent. You have chat with extra steps. Where does yours stall: missing knowledge, no stop condition, or no write-back?

2 0 3 34 0

View Details

Debo @DeepSData

3 days ago

Write-back is the one I see break first. The model finishes a loop with a guess, then the next loop treats that guess as fact. No source. No “still unverified” flag. So the loop looks busy and gets more wrong each round.

0 0 2 28 0

View Details

Debo @DeepSData

3 days ago

Share your domain and let me check how much it's worth.

0 0 3 21 0

View Details

Debo @DeepSData

3 days ago

@dev_maims Saving this. The later stages (agents + shipping) feel like the real jump. For someone who only has a few hours a week, would you compress Stages 1–6 or keep them sequential so the foundation doesn’t collapse?