AI coding agent cost controls at Replit, Kilo Code and Symbotic

Replit, Kilo Code and Symbotic described their AI coding agent cost controls at VB Transform 2026, as teams give agents more software work and face less predictable model bills. The common approach was to route tasks to cheaper models where possible, make usage visible to managers and retain human controls for work that needs review. The supplied report did not provide aggregate-spending figures for the companies.

The discussion came in a July 15 session on enterprise software and agentic coding featuring Kilo Code co-founder Emilie Schario, Symbotic distinguished engineer for AI and cloud Jared Go, and Replit head of product engineering Amol Jain. Their accounts were speaker-reported operating practices, rather than independently audited cost or productivity results.

How are companies controlling AI coding agent costs?

Kilo Code's Schario said companies can reserve costly frontier models for planning and architecture, then use less expensive open-weight models for later implementation work. Kilo Code supports more than 500 models through its gateway, she said, and model selection can also account for regional availability, data-retention requirements, isolated environments, keys and commits.

Schario said she is watching cost per pull request as a measure that links agent spend to delivered engineering work. She cited one internal engineer whose daily bill reached about $600, while arguing that the appropriate judgment depends on the output produced, not the bill alone. She also said customers have reported exhausting an annual AI budget accidentally, though no customer figures or underlying data were provided.

Symbotic sets monthly usage-cost tiers for employees, according to Go. Its internally built tool gives managers visibility into pull requests and usage patterns, and managers can move employees up or down those tiers. Go said a Cursor pricing change, which ended Symbotic's legacy flat per-request rate for frontier models, led employees to compare models for particular languages and coding tasks.

Replit uses routing and defaults intended to balance cost with capability, Jain said, because most tasks do not need a frontier model. The company also learned that agent spending can extend outside engineering: Jain described a support-side automation running on GPT 5.5 Pro Max that generated an unusually large bill. He did not disclose the amount.

What human controls remain in agentic coding?

Cost management is separate from software quality and access controls. Jain said Replit has an agent assess each pull request and assign a risk score. Authors can self-merge low-risk pull requests, while pull requests that are not low risk go to human reviewers for code review and feedback. He said Replit's agents run in cloud virtual machines with access controls behind token proxies.

Go said agents work well on greenfield projects, or entirely new codebases, but maintaining and changing existing systems remains harder. He also said agents do not make strong product decisions. The result is a practical constraint on autonomous coding: routing and budget caps can reduce spend, but they do not remove the need for judgment over what should be built or changed.

• Route planning and execution work to different models based on cost and capability.
• Give managers usage and pull-request visibility, with enforceable employee-level tiers.
• Measure spend against an engineering output, while keeping risk-based review and access limits.

This story draws on original reporting from VentureBeat.