The token cost debate keeps getting all the attention. The actual ROI line for autonomous agents will be the compute substrate: sandboxes, network bandwidth, retry overhead. Daytona is one of the first companies betting the deployment surface becomes the durable moat.
If most agents in 12 months are spawned, fail, and retried 10x to finish a task, the compute layer eats the margin before the LLM does. Anyone tracking per-agent infrastructure cost in production yet, or is it still hidden inside a generic cloud-bill line?
The retry tax is the part nobody's line-iteming yet. When Anthropic shipped Managed Agents on April 8, Fastly fell 18%, Akamai 16%, and Cloudflare 13% within three sessions, so markets are already pricing the substrate even while finance teams bury it in a generic cloud row. Daytona's moat holds only if that retry overhead stays legible enough to bill for.
The 0 to 50 percent RL/eval mix in a single quarter is the most informative number here. It is the AI infrastructure equivalent of a single workload type going from edge case to majority share at one customer cohort, and it explains the bare-metal-with-its-own-scheduler choice. The CLI-over-MCP framing is also the right call for agent ergonomics right now. CLI is composable, stable, well-documented, and works at every scale of compute, while MCP still has to grow into production traffic semantics and standardized error contracts.
Worth noting what pulled that RL/eval mix to half: TerminalBench, GDPVal, and Harbor all assume a computer, so the research side now grades agents on the exact toolkit the products ship. CLI wins partly because that's what the benchmarks already reward, and MCP wasn't in the room when those were written.
Most AI app companies do not want to become infra companies, but once they hit scale, they accidentally become one anyway. Their margins start depending on how efficiently they can schedule and isolate work, not just how good the model is. So the positive case is: Daytona lets AI companies scale labor-like products with software-like margins.
The token cost debate keeps getting all the attention. The actual ROI line for autonomous agents will be the compute substrate: sandboxes, network bandwidth, retry overhead. Daytona is one of the first companies betting the deployment surface becomes the durable moat.
If most agents in 12 months are spawned, fail, and retried 10x to finish a task, the compute layer eats the margin before the LLM does. Anyone tracking per-agent infrastructure cost in production yet, or is it still hidden inside a generic cloud-bill line?
The retry tax is the part nobody's line-iteming yet. When Anthropic shipped Managed Agents on April 8, Fastly fell 18%, Akamai 16%, and Cloudflare 13% within three sessions, so markets are already pricing the substrate even while finance teams bury it in a generic cloud row. Daytona's moat holds only if that retry overhead stays legible enough to bill for.
The 0 to 50 percent RL/eval mix in a single quarter is the most informative number here. It is the AI infrastructure equivalent of a single workload type going from edge case to majority share at one customer cohort, and it explains the bare-metal-with-its-own-scheduler choice. The CLI-over-MCP framing is also the right call for agent ergonomics right now. CLI is composable, stable, well-documented, and works at every scale of compute, while MCP still has to grow into production traffic semantics and standardized error contracts.
Worth noting what pulled that RL/eval mix to half: TerminalBench, GDPVal, and Harbor all assume a computer, so the research side now grades agents on the exact toolkit the products ship. CLI wins partly because that's what the benchmarks already reward, and MCP wasn't in the room when those were written.
Most AI app companies do not want to become infra companies, but once they hit scale, they accidentally become one anyway. Their margins start depending on how efficiently they can schedule and isolate work, not just how good the model is. So the positive case is: Daytona lets AI companies scale labor-like products with software-like margins.