The steering queue is the detail people skip. A model that can’t accept a correction safely mid-task isn’t very useful, no matter how well it benchmarks.
my local agent has trouble with tool calling... it simply refuses to do it most of the time. or it thinks it is too smart to lower itself to tool calling and prefers to give answer itself.... also want to know more about how to make agent create plan and then have it gradually implemented without getting stuck...
Awesome project! I’ve taken several deeplearning.ai courses, but nothing caught my eye for a follow-up project until I saw this. Thanks a lot for sharing this amazing content! Greetings from Durango, Mexico.
The clinical-AI translation of the harness point is that the model is only one part of the safety case.
For an OR-adjacent agent, I would want the harness to preserve source map, allowed tools, human checkpoints, replayable trace, sandbox boundary, and the state change it tried to make.
That turns "did the model answer?" into "can the team inspect what the agent touched before inheriting the output?"
The steering queue is the detail people skip. A model that can’t accept a correction safely mid-task isn’t very useful, no matter how well it benchmarks.
Exactly!
my local agent has trouble with tool calling... it simply refuses to do it most of the time. or it thinks it is too smart to lower itself to tool calling and prefers to give answer itself.... also want to know more about how to make agent create plan and then have it gradually implemented without getting stuck...
Fair and annoying issues! We will get into that in the 2nd lesson when building the agent loop. Which I will release next week
Thank you Paul.
My pleasure. Enjoy 🤩
It seems to be interesting.
I really like this one!
Awesome project! I’ve taken several deeplearning.ai courses, but nothing caught my eye for a follow-up project until I saw this. Thanks a lot for sharing this amazing content! Greetings from Durango, Mexico.
This means a lot to me, Edwin 🤩. Enjoy learning how to build an agent harness from scratch!
The clinical-AI translation of the harness point is that the model is only one part of the safety case.
For an OR-adjacent agent, I would want the harness to preserve source map, allowed tools, human checkpoints, replayable trace, sandbox boundary, and the state change it tried to make.
That turns "did the model answer?" into "can the team inspect what the agent touched before inheriting the output?"