AI Coding Agents Are Straining Budgets
How companies like Replit, Kilo Code, and Symbotic manage the costs
199% of Engineers No Longer Write Code
According to Kilo Code co-founder Emilie Schario, only 1% of engineers now write or read code—the rest is handled by AI agents. This shift raises new questions: which systems can be delegated, who fixes model errors, and whether rising token costs signify progress or budget waste. Companies are grappling with these challenges as AI becomes embedded in workflows.
2AI Excels in Greenfield, Struggles in Brownfield
Symbotic's Jared Go notes AI thrives in building new codebases (greenfield) but struggles with legacy systems (brownfield). AI handles heavy lifting, reducing the need for human code reviews. However, human intervention remains critical for product decisions. 'Agents don’t make strong product calls,' Go emphasizes, highlighting the balance between automation and oversight.
3Replit’s Human-on-the-Loop Approach
Replit takes a cautious stance: AI assigns risk scores to pull requests (PRs). Low-risk PRs auto-merge; high-risk ones go to human reviewers. 'It’s human on the loop, not in the loop,' says Amol Jain, Replit’s head of product engineering. Their internal tool acts like self-driving software: engineers assign tasks, and AI handles planning, implementation, and testing.
4AI Solves a 'Gnarly Bug' in 6 Hours
When a Replit engineer hit a dead end with a deep system bug, an AI manager agent took over. It spawned sub-agents to diagnose the issue and more to fix it. Six hours later, AI delivered a PR for a bug that stumped humans. Jain notes the agents run in secure cloud VMs with token proxies, ensuring safety while autonomously resolving complex issues.
5Multi-Model Flexibility with 500+ Options
Kilo Code’s gateway supports 500+ models, decoupling engineering tools from model choices. Schario advises using costly frontier models for architecture, then switching to cheaper open-weight models. Regional provider limits and data policies also influence routing. 'It’s about balancing cost, capability, and constraints,' she says, emphasizing tailored AI workflows for efficiency.
6Tokenmaxxing to Curb Runaway Costs
As AI adoption grows, so do expenses. Some firms cap usage via tokenmaxxing. Schario hears from clients who blew annual AI budgets accidentally. Kilo Code guides them to use expensive models for planning, cheaper ones for execution. Model Context Protocol (MCP) and skill-sharing further optimize output. 'Upleveling your team maximizes model ROI,' she adds.
7Cost-Per-PR as the Key Metric
Schario tracks cost per pull request: 'It’s the closest proxy for value.' A $600 daily bill may seem steep, but output can justify it. Symbotic tiers employee AI spending monthly, giving managers visibility into usage. 'Caps drive efficiency,' says Go. When Cursor ended legacy discounts, it forced a company-wide efficiency audit, underscoring cost-awareness.
8AI Redefines ROI for Enterprises
The issue isn’t spending—it’s spending without returns. AI reshapes ROI calculus, as efficiency gains offset costs. Symbotic’s tiered budgeting and Replit’s risk-scoring exemplify adaptive cost management. 'Spend with no ROI is the problem,' Schario concludes, highlighting how AI’s value lies not in cost reduction, but in transformative productivity.