The 1% Engineer and the Autonomous Infrastructure Shift
At Kilo Code, engineers are reading or writing code themselves only about 1% of the time, according to co-founder Emilie Schario at VB Transform 2026. The rest of the workflow is handled by agents. This paradigm shift forces hard questions onto engineering leadership. Teams must figure out which systems are safe to hand over to autonomous code generation, who cleans up when models goof up, how to support multi-model architectures, and whether skyrocketing API token bills represent actual progress or just a burned IT budget.
As far as tech leads from Replit, Kilo Code, and Symbotic are concerned, the trend is a welcome evolution. “Unless something’s really broken or debugging, 99% of the time engineers are not reading or writing code anymore,” Schario noted.
Greenfield Velocity Versus Brownfield Maintenance
Building brand-new codebases from scratch is remarkably frictionless for modern agents, but maintaining existing legacy architectures remains a persistent bottleneck. Jared Go, distinguished engineer for AI and cloud at warehouse automation company Symbotic, outlined strict criteria for deploying autonomous tools.
“These are my criteria,” Go explained. “Let’s look at it from the lens of security, elegance, clean, concise code, water tightness.” By enforcing these automated checks, the AI handles heavy lifting while minimizing the need for intensive human code review. Human intervention becomes necessary further down the line because agents struggle with high-level product decisions. “Greenfield is so easy for agents. Brownfield we all know is where the actual challenge lies,” Go stated.
Replit takes a different architectural tack. While the company has “gone very agentic,” product engineering head Amol Jain explained that they maintain a conservative posture toward automated code integration. Every pull request generated by an agent receives an automated risk score. Low-risk pull requests are self-merged by their authors, while high-risk changes route to human reviewers for code reading and feedback.
“The idea was human on the loop, not human in the loop,” Jain said. Replit runs an internal tool functioning as a self-driving system for software engineers. Developers assign tasks to agents that execute end-to-end planning, implementation, and testing. “It’s a fleet of agents that run in their own cloud virtual machines with access controls behind token proxies so they’re secure,” Jain added.
In one complex scenario shared by Jain, an engineer faced a persistent, gnarly bug deep within system architecture that defied conventional debugging. The task was handed to an AI manager agent, which instructed the system to run diagnostics overnight. The manager agent spun up underlying worker agents that located the anomaly, then deployed additional agents to implement a fix. Six hours later, the AI delivered a completed pull request for a bug that had baffled human developers.
Breaking Vendor Lock-In Through Multi-Model Gateways
Kilo Code’s gateway infrastructure supports over 500 distinct models, allowing engineering teams to decouple their agentic development workflows from underlying proprietary model APIs.
“Your software that you’re using to do agentic engineering should be decoupled from the model that you’re using to do it,” Schario said. Engineering teams frequently deploy expensive frontier-tier models for high-level architectural planning, then pivot to lower-cost open-weight models for routine implementation tasks.
Routing decisions also factor in localized data retention policies, secure computing perimeters, and specialized developer commits. “It’s factoring in what’s important to you, what limitations you’ve set, what data retention policies you’ve established, what keys you’ve brought in, what commits you might have … into that routing decision,” Schario noted.
Replit similarly manages a cost-versus-capability spectrum on behalf of its user base. “We are essentially making the decisions on users’ behalf of what model to use when, in what capacity, to minimize cost and maximize capability,” Jain said.
Controlling Runaway Token Bills and Establishing ROI
As enterprise adoption scales, runaway operational costs have forced organizations to track and cap consumption through token management strategies. Schario reports hearing from clients who admit, “I accidentally spent my whole AI budget for the year … so what do I do now?”
Kilo Code addresses this by guiding customers toward tiered workflows: deploying expensive proprietary models strictly for structural planning and utilizing open-weight alternatives for execution. Additionally, utilizing the Model Context Protocol (MCP) and sharing specialized internal skills helps teams maximize model output.
Schario pointed to internal engineers with a “heavy foot” who consistently top usage leaderboards, requiring periodic inquiries into their token consumption. While a $600 daily infrastructure bill initially triggers alarm, calculating total output often justifies the expenditure. “Cost per pull request is the metric that I’m paying attention to right now,” Schario said. “It feels like the closest proximity for how I can measure value.” Spend itself is not the issue; unreturned spend is.
Symbotic enforces hard monthly budget tiers for its employees through custom internal tooling that grants engineering managers clear visibility into pull requests and consumption trends. Managers adjust user tiers dynamically. “Having a cap and seeing how many people went up in cap this month makes a big difference when you’re trying to corral these costs and make things efficient,” Go said.
When Cursor ended a legacy discount that had grandfathered Symbotic into a flat per-request rate for frontier models, moving all users to standard enterprise pricing, it triggered an internal efficiency review. Developers immediately began experimenting with specialized models optimized for specific languages like C#.
Replit expanded autonomous agents beyond core software engineering into customer support operations, uncovering scenarios where non-technical staff rapidly depleted allowances. Jain noted that support agents running intensive tasks on advanced models like GPT 5.5 Pro Max quickly burned through capital. “At least till that point, the ROI was rather clear,” Jain said. “We could see engineering productivity 3X, so no one had questioned it yet.” Granular visibility, intelligent model routing, and sensible cost defaults remain vital. “Most tasks do not need the frontier.”
