AI Coding Agents Are Changing Software Economics Faster Than Companies Expected

Sapo

The most important change in software development may no longer be how fast engineers can write code.

It may be how little code they need to write themselves.

At Kilo Code, engineers reportedly spend only about 1% of their time reading or writing code, with AI agents handling the other 99% except when systems are broken or debugging is required. That sounds less like a productivity feature and more like a new production model for software companies.

But there is a catch.

When machines can generate enormous amounts of code, the scarce resource moves somewhere else. Companies now have to manage model selection, security, code quality, agent permissions, infrastructure and, increasingly, AI spending.

The question is no longer whether AI can write software.

The question is whether a company can build an economic system around machines that write software continuously.

The Main News

Technology leaders from Kilo Code, Replit and Symbotic describe a development environment where AI agents are moving from assistants to autonomous workers.

The change is particularly visible at Kilo Code. Its engineers reportedly spend almost all of their time directing agents rather than manually producing code.

Replit has taken a more controlled approach. Its internal system gives agents responsibility for planning, implementation and testing, while another agent evaluates pull requests and assigns risk scores. Low-risk changes can move forward with limited human intervention, while higher-risk changes receive human review.

Symbotic's experience highlights the other side of the equation.

AI is extremely effective when creating something new. It becomes more difficult when it has to understand an existing system full of historical decisions, technical debt and undocumented assumptions.

That is the difference between greenfield and brownfield development.

And it may become one of the defining constraints on enterprise AI.

What Is New?

The new development model is not simply "AI writes code."

It is agents execute software work from beginning to end.

At Replit, developers can give an agent a task and allow it to plan, implement and test the solution. The agents operate inside cloud virtual machines with access controls and token proxies.

That architecture creates something closer to a software production line.

One agent can manage the assignment.

Other agents can investigate.

Others can write code.

Others can test it.

Another system can evaluate the resulting pull request.

The human developer becomes less of a typist and more of an operator, reviewer and decision maker.

One example illustrates the potential.

A difficult bug that human engineers could not reproduce or solve was handed to an AI manager agent. The manager delegated work to multiple underlying agents, which investigated the problem and eventually generated a fix. Six hours later, an AI-generated pull request was ready.

That is a meaningful change in software economics.

The unit of work is moving from the engineer's keystrokes to the completed outcome.

Why It Matters

For decades, software companies scaled by hiring more engineers.

AI introduces another variable.

A company can potentially increase output without increasing engineering headcount at the same rate.

But that does not mean engineering becomes free.

It means the cost structure changes.

Instead of paying only for salaries, infrastructure and conventional developer tools, organizations increasingly pay for inference.

And inference is variable.

A simple request might require a relatively inexpensive model.

A complex architectural decision may require a frontier model, large context windows and multiple agent calls.

This creates a new management problem.

If AI becomes more productive but costs are allowed to grow without control, the company may simply be exchanging labor inefficiency for compute inefficiency.

That is why the next competitive advantage may be AI cost management.

How the Product Actually Works

The emerging architecture looks increasingly like a multi-agent system.

The developer defines the goal.

An agent decomposes the problem.

Other agents execute specialized tasks.

Tools provide access to repositories, documentation, infrastructure and other systems.

Testing agents validate the work.

A review system determines whether human intervention is necessary.

This is fundamentally different from asking a chatbot to generate a function.

The chatbot produces an answer.

The agent participates in a workflow.

That distinction matters because workflows can be measured.

A company can ask how many pull requests were created, how many were accepted, how long they took, how often humans intervened and what the final cost was.

That creates the foundation for real ROI analysis.

The Data and Infrastructure Layer

One of the most important developments is the move toward multi-model architectures.

Kilo Code's gateway supports more than 500 models. The company's argument is that the agentic development layer should be separated from the underlying model.

That separation is strategically important.

Models are becoming interchangeable components.

A company might use a powerful frontier model to design an architecture, then switch to a cheaper open-weight model for repetitive implementation.

The system can also take into account regional restrictions, data-retention policies, credentials, security requirements and isolated environments when choosing which model should execute a task.

This creates a new abstraction layer.

The user should not have to know which model is best for every task.

The orchestration system should decide.

The best AI coding platform may therefore become less like a model and more like a traffic controller for models.

The Economics of AI Coding

This is where the story becomes much more interesting.

AI spending can look frightening when measured in isolation.

An engineer generating a $600 daily AI bill sounds excessive.

But the correct question is not whether $600 is expensive.

The correct question is what $600 produced.

Kilo Code is therefore focusing on cost per pull request as a practical measure of AI value. The reasoning is simple. Spending itself is not necessarily the problem. Spending without measurable output is the problem.

That is a much better way to think about AI ROI.

Companies should stop asking only:

"How much are we spending on AI?"

They should also ask:

"How much useful work did that spending create?"

This changes the CFO's relationship with AI.

AI becomes closer to a variable production cost than a conventional software license.

Tokenmaxxing Has a Limit

Some enterprises have responded to rising AI costs by tracking usage and imposing limits.

Symbotic, for example, created monthly employee cost tiers and a management system showing pull requests and usage trends. Managers can adjust employees between tiers based on actual usage.

That is rational.

But excessive restrictions can create the wrong incentive.

If developers are afraid to use an agent because they might exceed a token allowance, the company has optimized the budget while damaging the productivity benefit it was trying to purchase.

The better solution is intelligent routing.

Use expensive intelligence where it creates disproportionate value.

Use inexpensive intelligence for routine work.

This is why the principle that "most tasks do not need the frontier" is so important.

The future is unlikely to be one model doing everything.

It will be a portfolio of models working at different price and capability points.

Platform or Product?

This is the strategic question.

An AI coding assistant is a product.

An orchestration layer connecting hundreds of models, repositories, tools, security policies and workflows is a platform.

The latter has considerably more strategic leverage.

If the model changes, the orchestration layer can remain.

If a cheaper model becomes better, the routing system can switch to it.

If a company changes its security requirements, the routing layer can enforce them.

This is similar to what happened across previous infrastructure cycles.

The application captures the user.

The platform captures the workflow.

And the orchestration layer can potentially capture the economics.

Competitive Advantage

The strongest advantage will not necessarily belong to whoever has the largest model.

It may belong to whoever understands which model should do which job.

Kilo Code's multi-model approach demonstrates one version of this strategy. Replit takes another approach by making model-selection decisions on behalf of users based on capability and cost.

That creates a new form of optimization.

Imagine a software organization where:

• Frontier models handle architecture.
• Smaller models handle repetitive implementation.
• Specialized models handle testing.
• Agents handle debugging.
• Human engineers handle product judgment.
• An orchestration layer manages security and cost.

That is not an AI assistant.

It is an AI-native software factory.

The Brownfield Problem

There is still a major weakness.

AI agents are excellent at greenfield development.

Existing software is harder.

Legacy applications contain undocumented dependencies, strange business rules, old APIs, inconsistent architecture and years of accumulated technical decisions.

Symbotic's Jared Go argues that this is where human involvement remains important because agents are not yet strong at making product-level decisions.

This creates an interesting paradox.

The more valuable an enterprise's software becomes, the more historical complexity it may contain.

And that complexity is precisely what makes autonomous coding harder.

The future of AI coding therefore will not be determined only by how well models generate new code.

It will depend on how well agents understand existing systems.

The Risk

There are three major risks.

The first is cost explosion.

Replit discovered that the problem was not limited to engineering. As agents expanded into other departments, a support employee generated an unexpectedly large bill because an automation was running on an expensive model.

The second is quality degradation.

Generating more code does not automatically mean generating better software.

The third is organizational blindness.

If employees no longer inspect code regularly, companies need better automated testing, security controls and review systems.

Otherwise, productivity gains can conceal technical risk.

The answer is not to put humans back into every step.

It is to put humans at the points where judgment actually matters.

That is the difference between human-in-the-loop and human-on-the-loop systems.

What Businesses Should Do

Companies should start by measuring outcomes rather than tokens.

Track:

Cost per pull request.

Time from task to production.

Human intervention rate.

Defect and rollback rates.

Model cost by task type.

Percentage of tasks using frontier models.

The second step is to build routing rules.

Do not allow every task to automatically use the most expensive model.

Use a stronger model when planning, reasoning or debugging requires it.

Use cheaper models for predictable implementation work.

The third step is to separate greenfield and brownfield workflows.

A fully autonomous system may be appropriate for isolated new projects.

Existing production systems require stronger permissions, testing and review.

Finally, companies should treat AI as a production system rather than another developer tool.

That means observability, budgets, access controls, audit trails and clear ownership.

The Broader Business Impact

The implications extend beyond software engineering.

Replit's experience shows that once agents move outside engineering, AI spending becomes an enterprise-wide management problem.

The same pattern will eventually appear in customer support, marketing, finance, operations and research.

Every department will be able to create automations.

Every automation will consume models.

And every model call will have a cost.

The result is a new enterprise architecture in which intelligence becomes a variable resource.

That makes AI routing as important as cloud infrastructure management once became.

The companies that understand this early will have an advantage.

Not because they necessarily spend less.

Because they will spend where the return is highest.

The Commerce Connection

There is a second AI transformation happening at the same time.

The software industry is changing how work gets produced.

Commerce is changing how products get discovered.

The supplied research reports that digital commerce journeys beginning on brand websites fell from 82% in 2014 to 38% in 2024, according to Salesforce research. At the same time, Adobe Analytics recorded more than 800% year-over-year growth in AI-driven traffic to retail sites.

That means the AI layer is increasingly sitting between companies and customers.

Consumers are asking AI what to buy.

They may receive a shortlist.

They may never visit a search engine.

They may never visit the brand website.

This creates a problem similar to the AI coding problem.

The old measurement system was designed for a different world.

For developers, traditional measurements focused on human engineering output.

For commerce, analytics traditionally focused on what happened after a visitor arrived.

AI moves the important decision earlier.

The New Invisible Funnel

A customer who visits a website and leaves without buying is measurable.

A customer who asks an AI assistant for recommendations, never sees your product and purchases from a competitor is much harder to measure.

There may be no abandoned cart.

No website session.

No referral.

No obvious lost conversion.

That absence is becoming strategically important.

The supplied research cites a 2025 zero-click study in which 60% of searches ended without a click. It also reports research involving 1,500 US consumers in January 2025, finding that a majority of AI-assisted product researchers made purchasing decisions directly from AI recommendations without returning to a search engine or brand website.

For brands, this changes the definition of visibility.

Ranking on a search page is no longer enough.

A company increasingly needs to understand how AI systems describe it, whether those systems recommend it and which competitors appear alongside it.

What Businesses Should Do Next

The same principle applies to both software and commerce.

Measure the new layer.

For software teams, measure the economics of agent execution.

For brands, measure AI discoverability.

Ask:

Does AI mention our company?

Does it recommend our products?

What competitors appear in the same answer?

How does AI describe our positioning?

Are those descriptions accurate?

What happens when consumers ask for alternatives?

These questions are becoming infrastructure questions, not simply marketing questions. The source material argues that existing analytics systems were not designed to measure the journey from consumer intent to AI-mediated discovery.

That is an important distinction.

The old internet was optimized around clicks.

The emerging internet is increasingly optimized around answers.

Conclusion

The biggest mistake companies can make is treating agentic AI as another software feature.

It is not.

AI coding agents are changing the production function of software.

AI answer engines are changing the distribution function of commerce.

In both cases, the old measurement systems become incomplete.

For software, counting engineer hours becomes less useful when agents perform most of the implementation.

For commerce, counting website traffic becomes less useful when AI determines the shortlist before the consumer arrives.

The winners will therefore build systems around the new units of value.

For software, that may be useful output per dollar of inference.

For commerce, it may be AI visibility and recommendation share.

The technology will continue to improve.

Models will become cheaper. Agents will become more capable. Routing will become more sophisticated.

But the strategic lesson is already visible.

Intelligence is becoming abundant. Good orchestration is becoming scarce.

The companies that learn how to allocate AI capability, control its cost, measure its output and integrate it into real workflows will have a much better chance of turning the AI boom into durable economic value.

The goal is not maximum AI usage.

The goal is maximum useful work per unit of AI spend.

The AI Coding Boom Has a Token Bill Nobody Can Ignore

SEO keywords

AI coding agents, agentic coding, AI software development, Replit AI agents, Kilo Code, AI coding costs, cost per pull request, multi model AI, AI model routing, enterprise AI agents, AI developer productivity, brownfield software, greenfield development, AI inference costs, AI orchestration