At Kilo Code, engineers are studying or writing code themselves solely about 1% of the time now, in accordance with co-founder Emilie Schario — the remaining is brokers. That shift is forcing new questions onto dev groups: which programs are protected handy over, who cleans up when fashions goof up, easy methods to assist multi-model architectures, and whether or not skyrocketing token payments imply actual progress or simply burned IT price range.
So far as tech leads from Replit, Kilo Code, and Symbotic are involved, it’s a pure — and welcome — evolution as agentic AI turns into embedded into increasingly enterprise workflows.
“Except one thing's actually damaged or debugging, 99% of the time engineers usually are not studying or writing code anymore,” Emilie Schario, co-founder of Kilo Code, stated at VB Rework 2026.
AI good at greenfield, not so nice at brownfield
For Jared Go, distinguished engineer for AI and cloud at warehouse automation firm Symbotic, the present second is about directing the main target of AI. "These are my standards," he stated. "Let's have a look at it from the lens of safety, class, clear, concise code, water tightness." That manner, AI does many of the heavy lifting, and human code evaluate isn't as essential.
Human involvement turns into obligatory additional down the road, Go famous, as a result of brokers don't make robust product selections. “Greenfield [building brand new codebases] is very easy for brokers. Brownfield [writing, updating, or maintaining existing code] everyone knows is the place the precise problem lies.”
Replit takes a little bit of a special tack: Whereas the corporate has "gone very agentic," they've been extra conservative with AI coding, defined Amol Jain, head of product engineering. An agent critiques every pull request (PR) and assigns it a threat rating; low-risk PRs are self-merged by their writer, whereas others go to human reviewers who learn the code and provides suggestions.
“The thought was human on the loop, not human within the loop,” Jain stated. Replit’s inner instrument is basically self-driving for software program engineers; devs give a job to brokers, which do finish to finish planning, implementation, and testing.
“It's a fleet of brokers that run in their very own cloud digital machines (VMs) with entry controls behind token proxies in order that they're safe,” Jain stated.
He shared one instance the place an engineer couldn’t repro or remedy a “very gnarly bug” deep in its programs. It was despatched to an AI supervisor agent, which informed it to fall asleep. The supervisor agent then spun up a bunch of underlying brokers that discovered the problem; it subsequently spun up a bunch extra brokers that discovered the repair. Six hours later, AI had a PR prepared for the bug that had puzzled human engineers.
Multi-model is the long run
AI suppliers are additionally evolving past the lock-in mannequin, as clients more and more demand multi-model alternative.
Kilo Code, for its half, helps 500-plus fashions in its gateway. "Your software program that you just're utilizing to do agentic engineering needs to be decoupled from the mannequin that you just're utilizing to do it," Schario stated.
As an illustration, Schario stated firms usually use costly frontier-tier fashions to architect a mission, then swap to a cheaper open-weight mannequin for the remainder of the work.
It’s additionally necessary to respect mannequin supplier limitations, resembling when they should work in closed or remoted environments or suppliers of their particular areas. “It's factoring in what's necessary to you, what limitations you've set, what knowledge retention insurance policies you've established, what keys you've introduced in, what commits you might need … into that routing resolution,” Schario stated.
Replit, equally, tends to have a greater sense of the associated fee versus functionality spectrum than its clients, Jain contended. “We’re primarily making the selections on customers' behalf of what mannequin to make use of when, in what capability, to attenuate price and maximize functionality.”
To tokenmaxx or to not tokenmaxx
After all, an necessary consideration as AI adoption will increase is runaway prices, which has led to some enterprises monitoring and capping AI use by tokenmaxxing.
Issues come from either side, Schario stated: internally and from clients. From the latter, she's listening to, "I by chance spent my complete AI price range for the 12 months … so what do I do now?" In response, Schario stated Kilo Code factors clients to the identical workflow: use costly fashions for planning, then open-weight fashions for affordability.
Additional, sharing abilities, robust steering, and Mannequin Context Protocol (MCP) will empower fashions. “Realizing the place you may actually uplevel your staff to assist them get essentially the most out of the fashions they're utilizing goes to make an enormous distinction,” Schario stated.
Internally, in the meantime, Schario famous one explicit engineer that has a "heavy foot" and is continually on the high of the utilization board. "I repeatedly must nudge, 'What are you doing there?'" she stated. It's straightforward to take a look at a $600 invoice for day by day work and react, "Wow, that's a lot," however trying on the quantity of labor accomplished can generally justify the associated fee.
“Price per pull request is the metric that I'm listening to proper now,” Schario stated. “It feels just like the closest proximity for the way I can measure worth.” Finally, AI adjustments how enterprises are desirous about ROI as a result of spend shouldn’t be the issue. “The spend with no return on that spend is the issue.”
Symbotic, for its half, has set per-month price tiers for its workers. The corporate constructed a instrument that offers managers visibility into PRs and utilization tendencies. They’ll then transfer customers up or down a tier as they see match, Go defined. “Having a cap and seeing how many individuals went up in cap this month makes an enormous distinction if you're attempting to corral these prices and make issues environment friendly,” Go stated.
When Cursor — which Symbotic makes use of closely — ended a legacy low cost that had grandfathered the corporate right into a flat per-request fee even for frontier fashions, and moved everybody to full pricing, it compelled a company-wide counting on effectivity, Go stated. "Folks have been saying, 'You need to do that mannequin … This works higher for this C# code, this no matter,'" he stated.
However the associated fee drawback is more and more transferring out of IT; Replit, for one, broadened brokers past engineering, and ultimately discovered {that a} person on the assist aspect had "blown by an insane sum of money," Jain stated. After they seemed underneath the hood, they found out it was as a result of they have been working an automation on GPT 5.5 Professional Max.
“Not less than until that time, the ROI was slightly clear,” Jain stated. “We might see engineering productiveness 3X, so nobody had questioned it but.”
Visibility that isn’t “anti-productive,” mannequin routing, and wise defaults are essential, he emphasised. “Most duties don’t want the frontier.”
