Technical leads from Replit, Kilo Code and Symbotic say agent-driven AI is handling the bulk of coding work in their organizations, bringing productivity gains but also new operational and cost-management challenges. At VB Transform 2026 they described differing approaches to safety, model selection, visibility and spending controls as enterprises embed agentic workflows into more parts of the business.
How agentic is development today?
Emilie Schario, co-founder of Kilo Code, said their measurements show engineers now read or write code themselves only about 1% of the time — the rest is done by agents. That shift creates new responsibilities for dev teams: which systems are safe to hand over to agents, who cleans up when models err, how to support multi-model architectures, and how to respond to soaring token bills.
Greenfield versus brownfield work
Jared Go, distinguished engineer for AI and cloud at Symbotic, emphasized that agents excel at greenfield work — creating new codebases — but struggle with brownfield tasks such as modifying or maintaining existing systems. Because agents do not make strong product decisions, human involvement becomes necessary later in the lifecycle when product and architectural judgment matters.
Replit’s "human on the loop" approach to PR risk
Replit has taken a more cautious stance on AI coding despite becoming "very agentic," according to Amol Jain, head of product engineering. Their internal tooling has an agent review each pull request (PR) and assign a risk score. Low-risk PRs can be self-merged by the author; higher-risk PRs go to human reviewers for code reading and feedback. Jain framed this as “human on the loop, not human in the loop.”
Replit runs a fleet of agents in dedicated cloud VMs with access controls behind token proxies. Jain shared an example where a hard-to-reproduce bug was assigned to an AI manager agent that spun up subordinate agents to investigate and then further agents to produce a fix. Six hours later the AI had produced a PR for the issue that had puzzled human engineers.
Multi-model strategies and decoupling models from tooling
Providers and customers are moving away from single-model lock-in toward multi-model choice. Kilo Code’s gateway supports more than 500 models, Schario said, arguing that the software used for agentic engineering should be decoupled from the underlying model. A common pattern is to use expensive frontier-tier models for initial architecture and planning, then switch to cheaper open-weight models for implementation.
Model routing must also consider provider limitations, such as the need to run in closed or isolated environments or comply with region-specific requirements. Routing decisions should factor in what matters to the customer: data retention policies, keys in use, commit histories and other constraints, Schario added.
Replit similarly claims they often make the decision on users’ behalf about which model to use when, to minimize cost and maximize capability.
Token bills, tokenmaxxing and cost controls
Rapidly rising token costs have pushed enterprises to track and cap AI usage. Schario hears concerns both internally and from customers — for example, customers saying they accidentally used their entire annual AI budget. Kilo Code suggests using expensive models for planning and cheaper models for execution.
Schario highlighted the importance of sharing skills, strong guidance, and using the Model Context Protocol (MCP) to help teams get the most out of models. She watches the “cost per pull request” metric as a proximate measure of value: spending is not the problem, spending with no return is.
Symbotic implemented per-month cost tiers for employees and built tooling that gives managers visibility into PRs and usage trends, allowing them to move users between tiers as needed. Jared Go said that seeing caps and how many people moved up in a month helps corral costs.
When Cursor, heavily used by Symbotic, ended a legacy discount and moved customers to full pricing for frontier models, that forced Symbotic to examine efficiency and model choice.
At Replit, costs also migrated out of IT control when agent use expanded beyond engineering. Jain recounted finding that a support-side user had consumed a huge amount by running an automation on GPT 5.5 Pro Max. Up to that point, ROI had been clear — Replit observed roughly 3X engineering productivity — so there had been little pushback on the expense.
Operational lessons: visibility, model routing and sensible defaults
All three companies emphasized the need for non‑disruptive visibility into usage, sensible model routing, and reasonable defaults because most tasks do not require frontier models. Multi-model support, secure agent environments, and clear cost metrics help organizations harness agentic AI without letting costs spiral.
Key open questions remain: which tasks should be delegated to agents, how to ensure security and compliance, who bears responsibility for model errors, and how to measure true ROI from these new workflows.
When these observations were made
The comments and examples were shared by Replit, Kilo Code and Symbotic representatives during a session at VB Transform 2026.
