Build Self-Learning AI Agents With Model, Harness, Context And User Feedback

@svpino The only way an agent can take feedback from it's users is by using said feedback to train a new version of the model, OR train some sort of bootstrap model who's only job is to take the input prompt and refine it with the feedback it was trained on.

@svpino 第三点被时间线吃掉了吗

@svpino most people just prompt and pray

@svpino @threadreaderapp unroll

@svpino the learn from user corrections point is the one most teams never get to because they're still arguing about which model to use.

The point that nobody watches the user might be the most underrated idea in agent design right now. Everyone obsesses over the model and harness, but the moment a user corrects the agent is the richest signal you'll ever get, and most systems just throw it away. This reframed how I think about my own builds. Thank you for laying it out so clearly.

@svpino cool

@svpino Sounds like a solid roadmap! Learning from various areas is key to growth. Excited to see where this journey leads!

@svpino training with real human interactions and decisions Nice

@svpino Balancing model development with practical user interactions truly makes a difference.

@svpino the model layer is the one everyone fixates on and can't move. harness and context are the only levers you own in prod

@svpino the user-feedback loop is the closest thing to an ltv moat in the agent space. every correction is a labeled data point that your competitor's agent doesn't have

@svpino the honest answer is that most people get tripped up in the harness, thinking the model is where the real learning happens, but that's just the easy part. the hard part is figuring out how to connect the model's output to real-world actions without breaking everything.

@svpino Appreciate it

@svpino developing agents that improve over time is probably one of the biggest jumps in AI this year

@YannRibemont @svpino @YannRibemont Hello, the unroll you asked for: https://threadreaderapp.com/thread/2070210421995569537.html Have a good day. 🤖

@svpino The harness is where most teams can actually improve the agent.

The model is mostly someone else's roadmap.

But the checks, tools, task boundaries, feedback capture, and recovery path are yours. That's where compounding starts.

@svpino The harness learning loop is clean in theory until the safety layer's false positives cost more than any model improvement.

@svpino Learning from every time a user corrects it is such an obvious signal but barely anyone builds for it.