Teams with agents in production are chasing the optimal agent. Usually they pick one lever: tune the harness (the prompts, tools, memory, and workflow around the model), or fine-tune the model, but the two depend on each other. A fine-tuned model was trained around the old harness, and a tuned harness was shaped around the old model, so whichever you change last leaves the other out of date. The answer is to optimize them together with model-harness co-training, an emerging field in AI agents.

In this talk, we’ll share what two years of building model-harness co-training have taught us about its core algorithmic questions. Should the model and harness learn from the same cases, take turns with alternating learning steps, or run as a nested optimization loop? How do you keep what each one learns from diverging or contradicting the other? When should a lesson that starts in the harness be distilled into the model’s weights? And with a fixed compute budget, how much should go to optimizing the harness
versus training the model?

Main Takeaways:
1. Why an agent’s model and harness should be optimized as one system
rather than as two separate projects.
2.The key design decisions involved in co-training, and the trade-offs behind
each.