99% of Kilo Code’s engineering is now agents — here’s the cost problem
At Kilo Code, engineers are writing or reading code themselves just 1% of the time. The other 99% is handled by agents, according to co-founder Emilie Schario. That’s not a future projection — it’s current operating reality, shared at VB Transform 2026.
The same shift is underway at Replit and warehouse automation company Symbotic. And all three are wrestling with the same side effect: AI that works this hard burns serious money.
Agents Are Running the Show
Replit has “gone very agentic” but built a risk layer on top, according to Amol Jain, head of product engineering. An agent scores every pull request. Low-risk PRs get self-merged. Higher-risk ones go to a human reviewer. Jain calls the posture “human on the loop, not human in the loop.”
Replit’s internal setup runs a fleet of agents in isolated cloud VMs behind token proxies. Jain shared one case where a gnarly bug had stumped human engineers. They handed it to an AI manager agent, which spun up a set of sub-agents to find the issue, then spun up more agents to find the fix. Six hours later there was a PR ready to merge.
For Symbotic, Jared Go frames agent use around criteria: security, elegance, clean and concise code, and water tightness. His observation on scope is worth noting: greenfield work is easy for agents. Brownfield — maintaining or updating existing codebases — is where things get hard.
The Cost Problem Is Real
Token costs are the pressure point nobody budgeted for. Schario described customers who have “accidentally spent their whole AI budget for the year.” Her prescription: use expensive frontier models for planning and architecture, then switch to cheaper open-weight models for execution.
Internally at Kilo Code, one engineer consistently tops the usage board with a $600 daily bill. Schario doesn’t automatically treat that as waste. “Cost per pull request is the metric I’m paying attention to right now,” she said. “It feels like the closest proximity for how I can measure value.”
Symbotic built internal tooling that gives managers visibility into per-employee usage and PR trends. Employees are assigned monthly cost tiers and can be moved up or down. When Cursor ended a legacy flat-rate discount and moved Symbotic to full pricing on frontier models, it forced a company-wide conversation about which model actually fits which task — including which works best for C# code specifically.
Replit found the cost problem spreading beyond engineering entirely. A support-side employee ran an automation on GPT-5.5 Pro Max and burned through a significant sum. Up to that point, Jain said, no one had questioned AI spend because engineering productivity had been tracking at 3x. Visibility, sensible model routing, and sane defaults are now priorities. “Most tasks do not need the frontier,” Jain said.
The Takeaway for Operators
The emerging management pattern across all three companies: decouple your tooling from any single model, build visibility into who is spending what, cap by tier rather than by fear, and measure output (PRs shipped, bugs fixed) rather than input (tokens consumed). Schario put it plainly — spend is not the problem. Spend with no return is.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
