Cost control in the era of Agentic Development

When AI-powered developers and AI agents do the work, the cost curve changes shape. Compute spikes are no longer bound by how many humans are at keyboards; they're driven by how many agents are running in parallel, how much context each one loads, how many retries it takes to produce code worth keeping, and how often a rollback sends the whole loop back to the start. At the same time, cloud spend explodes as new AI infrastructure arrives with increasingly complex and expensive demands. Token spend, evaluator chains, sandboxing strategies, and context budgets, none of which used to sit anywhere near the FinOps conversation, are now urgent problems to solve. A team shipping 10,000 PRs a day with 500 developers is operating on a fundamentally different cost curve than one shipping 1,000 PRs with 1,000 developers, and the levers that govern that curve are still being figured out in production. This roundtable is for platform leaders already in the middle of it. Where is your spend actually going if agents are generating most of it, what guardrails have held up under real workloads, how are you balancing context-richness against the bill, and what are you building next to keep the curve from running away from you?