AI coding agents are blowing through budgets — Replit, Kilo Code, and Symbotic explain how they're managing it

At Kilo Code, engineers now read or write code themselves only about 1 percent of the time, according to co-founder Emilie Schario; agents handle the rest. That figure is the concrete signal behind a broader industry conversation VentureBeat covered with tech leads from Replit, Kilo Code, and Symbotic: AI coding agents are not a side experiment anymore, and the cost of running them is showing up as real budget pressure.

The technical shift is less about a single product and more about how work is split. When agents own nearly all of the code path, teams have to decide which systems are safe to hand over, who cleans up when models goof up, and how to support multi-model architectures instead of one fixed model. Those are operational design choices: boundaries of agency, failure ownership, and routing across models that may differ in cost, capability, and failure modes.

For engineers and builders, the 1 percent human-coding figure reframes day-to-day work. Review, verification, and repair become the primary human jobs when agents produce most of the diffs. That changes how teams staff reviews, define safe handoff points, and decide when a human must still touch the code. Token spend becomes part of the same conversation as velocity: progress has to be measured against the bill, not only against output volume.

The market context is that Replit, Kilo Code, and Symbotic are all living with this tradeoff at once. Different products and domains, same pressure: agents scale coding throughput while multi-model setups and high token usage can empty IT budgets if no one is checking whether the spend tracks useful work. The shared stance from those leads is that this tension is natural for teams that have pushed agents into the center of development.

The practical takeaway is to treat agent spend as a first-class engineering metric. Track which systems agents may touch, assign clear cleanup ownership when models fail, and design multi-model routing so cost and risk are explicit. Watch whether token bills stay tied to measurable gains in shipped work or drift into pure burn with little extra progress.