Quick Facts
- Agentic coding tasks consume approximately 3,500 times more tokens than single-round code reasoning, driven by compounding input token accumulation.
- Replit engineers nearly tripled code output over six months while keeping quality metrics flat and reversions stable.
- Gartner forecasts AI agent software spending will reach $207 billion in 2026, up 139% from 2025, while 62% of organizations cannot predict their monthly AI expenses.
AI coding agents are delivering real productivity gains. They are also torching budgets at a pace that caught even the largest technology companies off guard. At VB Transform 2026, engineering leaders from Replit, Kilo Code, and Symbotic laid out how they spotted the problem and what they built to contain it.
The scale of the spending shock is real. Uber burned through its entire 2026 AI coding budget by April. Microsoft pulled Claude Code licenses from developers months after rolling them out. A Priceline employee reported a routine Cursor renewal came back four to five times more expensive than expected.
Why Agents Cost So Much
The core issue is architecture. In a standard code reasoning task, a model reads a prompt, reasons, and outputs an answer. An agent does something different: it reads the task, gets a response, then re-reads everything before the next action, then re-reads all of that plus the new response before the next action. The context window balloons with every step.
Research cited at the conference found that agentic tasks consume roughly 3,500 times more tokens than single-round reasoning and 1,200 times more than multi-round chat, with input tokens driving most of the cost. Token usage also varies widely: runs on the same task can differ by as much as 30 times in total token count, and higher token use does not produce higher accuracy.
Ramp’s June 2026 AI Index, which covers more than 70,000 U.S. businesses, shows the spending range has become extreme. The top 1% of companies by AI spend per employee averaged $7,449 per employee per month. The median was $11.38.
How Three Companies Are Responding
Emilie Schario, co-founder of Kilo Code, said engineers at her company now write or read code themselves only about 1% of the time. Agents handle the rest. Kilo Code routes tasks across four model tiers: Efficient, Frontier, Balanced, or Free. Its Auto Model feature handles routing automatically based on a session’s budget and the complexity of the work, steering spend toward cheaper models when the task allows it.
Jared Go, distinguished engineer for AI and cloud at Symbotic, described a per-month cost tier system built for its workforce. Symbotic gives managers a dashboard showing individual pull request counts and usage trends. Managers can move employees up or down a tier based on actual output. Go said visibility alone changed behavior. When Cursor ended a legacy flat-rate discount and moved the company to full pricing, the team was forced to compare models by task type. Engineers started sharing which models handled specific languages more efficiently.
Replit’s approach centers on risk scoring. An agent reviews every pull request and assigns a risk score. Low-risk PRs are self-merged by their authors. Higher-risk ones go to human reviewers. Amol Jain, head of product engineering at Replit, described the design as “human on the loop, not human in the loop.”
Replit agents now handle end-to-end planning, implementation, and testing. They also investigate production incidents, triage support tickets, research sales accounts, and analyze business data. Each agent runs in its own cloud virtual machine with access controls behind token proxies. Keeping a consistent cohort of engineers, Replit measured 2.9 times as much code output over six months, with review times, reversions, and product incidents holding flat.
The ROI case held until it didn’t. Jain said a support-side employee outside engineering ran an automation on GPT 5.5 Pro Max and “blew through an insane amount of money.” The engineering team had not been watching non-engineering usage. “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.”
What This Means for Software Leaders
The three companies share a common approach: set hard spending tiers, build visibility tools for managers, and route tasks to cheaper models when accuracy requirements allow it. Waiting for invoices to surface the problem is too slow. By the time the bill arrives, the budget is already gone.
Uber CTO Praveen Neppalli Naga put it plainly: “I’m back to the drawing board, because the budget I thought I would need is blown away already.” For any software company scaling agentic workflows, that sentence is the benchmark to beat.