At Kilo Code, engineers read or write code themselves only 1% of the time. The other 99% is agents, according to co-founder Emilie Schario, speaking at VB Transform 2026. Replit and Symbotic are in similar territory: Replit deploys a fleet of agents running in isolated cloud VMs behind token proxies, each PR scored for risk and self-merged or escalated accordingly. One gnarly bug that stumped human engineers was resolved in six hours by a manager agent that spun up sub-agents to find the issue, then more sub-agents to find the fix.
The technical fault lines are specific. Symbotic distinguished engineer Jared Go flags greenfield work as easy for agents, brownfield as the real problem. His review criteria: security, elegance, conciseness, water tightness. Kilo Code routes across 500-plus models, using expensive frontier models for architecture and cheaper open-weight models for execution. Replit makes model selection decisions on behalf of users to balance cost and capability. The read worth having is in how each company draws the line between human-on-the-loop and human-in-the-loop, and why those lines are different.
Cost is the pressure point nobody has fully solved. Symbotic caps users by monthly tier, built internal tooling for manager visibility into PR output versus spend, and got a company-wide efficiency reckoning when Cursor ended a legacy flat-rate discount on frontier models. Replit found a support-side employee burning through budget running automations on GPT 5.5 Pro Max. Schario's working metric is cost per pull request. Her framing is blunt: spend is not the problem, spend with no return is. Most tasks do not need the frontier.
[READ ORIGINAL →]