Why it matters
These three problems (isolating what an agent can touch, how fast tokens actually get served, and whether a change made the agent better or just different) are exactly the layers that determine your real cost per successful task, not just cost per token.
The tokenmaxxing angle
Fireworks pitches serving fine-tuned open models at agent-loop speed as an alternative to always calling a frontier API; stacking that against Braintrust's evals is how you'd actually prove a cheaper model swap didn't quietly tank quality.
From the organizers
The listing gives the format as lightning talks from E2B, Fireworks AI, and Braintrust in SoMa, with each company’s one-line pitch (E2B: isolated machines for agents; Fireworks: turning open models into owned, tuned production models).