Vinit Agrawal, co-founder and CTO at Tars, takes the definition he likes best and builds the whole video on it. An AI agent is the model plus the harness. The model is the brain. The harness is everything else, and everything else is the part you control. A handful of labs build the models and ship them to you. What you do around the model is yours, and that is where the difference between a useful agent and a text machine gets made.

From there it is a working explanation of what a harness is made of. A raw model with nothing around it does not remember the question you asked two seconds ago, so the memory you experience in a chat product is the harness at work. Context has to be chosen and passed in. Tools have to be described, requested, and executed by something that is not the model. The loop around all of that has to know when to stop. He closes on his own experience of building and rebuilding this at Tars for a customer experience product, and on the point worth carrying into your next model release: swapping the model is a config change, upgrading the harness is the work.

What this video covers

• Why an agent is the model plus the harness, and why the harness is the only part you control
• What a raw model does when you ask it a follow-up question, and why chat products seem to remember anyway
• The four kinds of context a harness feeds the model, instructions, knowledge, memory, and skills
• How a tool call actually runs, with the model requesting and the harness executing
• The agent loop, and the two things a model has to get right for it to finish
• Why the same model with a different set of tools becomes a completely different product

Chapters

• 0:00 Why a harness is getting more attention than a model
• 0:45 What a harness actually is
• 3:18 The horse and the motor
• 4:56 The model forgets
• 6:29 Context
• 8:20 Tools and the agent loop
• 11:03 The shift from how to what
• 12:08 Upgrade the harness