More code does not equate to more outcomes, and the two are being quietly confused across many enterprises as leaders strive to adopt AI-assisted development. This confusion is expensive and challenging the ROI across many IT departments.
Every CIO is under pressure to prove efficiency gains and move toward agentic engineering. Every developer feels the same pressure: use the agents, ship faster, and fundamentally, do more with less. The tooling is real, the models are impressive, and the agents give hope to what was previously an impossible task.
Today, anyone can take a prompt to a working application in minutes. But behind the guise of a “working application” there are 1000s of lines of opaque code. Built with insecure patterns, redundant loops and architectural decisions that no one can explain. As AI-assisted development continues to democratize software development, the skillset to deeply understand and interrogate the decisions being made by agentic harnesses becomes crucial to building at enterprise grade.
The hidden cost
Independent research has found that AI-generated code introduces roughly 1.7X more defects than human-written code across logic, readability, security, and error handling.¹ Veracode's testing of more than 100 models put the multiplier on serious security vulnerabilities at 2.74X,² demonstrating that generating code faster is not a productivity gain. Enterprises are also seeing a significant rise in “code churn,” which has doubled in the last 12 months. This occurs when code is re-written or discarded within two weeks of its initial development. Further highlighting that developers are spending more time remediating their AI outputs than developing new features.3
Without the appropriate enterprise structures, AI-driven development is a cost that compounds during the lifetime of the application across bug fixes, maintenance, and incident management. An agent with too much freedom becomes an application, which is a liability that organizations must test, secure, and maintain for years.
When we step back and strip away the hype, the core principles of software development are exactly what they have always been. Enterprises need applications that are secure, performant, high quality, and relevant to the business problem in front of them. They need applications that can be scaled instantly to thousands of employees and millions of customers to deliver real enterprise value.
Engineering the boundaries of AI agents
I know that we can all see that the promise of agentic engineering is there. As we have experimented, understood the pitfalls of LLMs and AI systems, AI leaders have evolved to recognize that they need structure in their agents, and this brought about prompt engineering – the ability to provide agents with better instructions, tested instructions where agents were told what to do, and how to execute tasks. Then we realized that prompts are not enough, so context engineering was identified as a discipline – the ability to give your agent context, refine context and ultimately provide scope along with the prompt instructions to give structure to execution. These are two disciplines every AI-assisted development team is investing in, and both do useful work, but neither close the gap. They share a blind spot that the agent should be free to act on instruction alone.
At enterprise scale, this freedom comes at a cost, because a capable agent with a good prompt and rich context will still improvise its way into insecure patterns, redundant code, and drifted intent.
The missing discipline is harness engineering, designing the environment the agent operates inside. A harness is the set of guardrails, ordered steps, checks between them, definition of success, and boundaries the agent cannot cross. Prompt engineering asks what to say. Context engineering asks what the agent knows. Harness engineering asks the more consequential question of what the agent is allowed to do, in what order, and how each step gets verified. That is the question that determines whether output is enterprise-grade, and the one leader will spend the least on.
Harness engineering in practice: Pega harnesses your coding agents
A harness is a set of enforced design choices. Executed together, it produces predictable enterprise-grade outcomes with coding agents:
The foundation of the Pega Platform™ is a model-driven architecture, which changes the unit of work. Instead of asking an agent to author and maintain thousands of lines of bespoke code, it configures a model that already carries the platform's scalability, security, and performance guarantees. Less code to write means less to review, less to get wrong, and far less to maintain. A model-driven application is enterprise-grade by construction, rather than made enterprise-grade after the fact by bolting on the frameworks, governance, and controls that java development forces you to build before anything reaches production.
Inside that assisted AI architecture, prescriptive authoring governs how the agent moves. It replaces the single open-ended prompt with a guide to understand the intent, align, plan, execute, and review. Each stage constrains the next, so the agent never guesses what "done" means and never runs ahead of a plan.
Where the Dev task is predictable, deterministic workflows remove the agent from the build entirely. The agent identifies the intent, and when the task is known – such as building a REST API or generating a “connection Gen AI” rule – there is no reason to let a probabilistic model reinvent it every time. It applies best practice identically every time, narrowing the scope where agent reasoning is required.
Every AI-assisted execution is bound through branched development to isolate the changes. Multi-developer or multi-agent work is isolated in its own branch, tested independently, and verified before it’s merged. Branching confirms the right thing was built and contains the blast radius of any issues.
Where developers want flexibility, Pega’s 100% open agentic architecture allows every developer and IT leaders to choose their models, choose their enterprise coding agents, and choose the systems that they interact with.
Enterprise grade by design
Utilizing the coding agent and platform harness in this unique way is what makes AI-assisted development with Pega enterprise grade. Inside Pega Infinity Studio, the AI-first development environment, every developer can take their re-imaged agentic processes and generate an implementation plan for their coding agent to execute. And when they press go, the application is built with security, governance, and explainability within the harness.
The harness allows every enterprise to unlock speed and certainty in agentic development, this not only closes the gap between prototype and enterprise grade, but it drives better results at the same speed of native coding agents, all at a predictable cost. There is zero token cost when you build with Pega, allowing developers to drive predictable outcomes and ROI for their companies.
Pega’s approach to open agentic engineering unlocks enterprise software development to any coding agent. While it is available natively through the AI-first development environment, Infinity Studio, the Infinity Build MCP toolkit powering agentic software development is available to any external coding agent. Developers can utilize their enterprising coding agent of choice, while still inheriting Pega's guardrails and model-driven architecture to develop code with confidence. The agents move fast, inside the harness that produces reliable, model-driven, enterprise-grade applications.
Driving scalable ROI with open agentic engineering
Where the tooling and model market is the loudest, it’s easy to focus on investing in better prompts and richer context. However, the sequence of execution remains the downfall to any “instruction-led development” as agents will continue to improvise against the plan and build from the ground up every time.
Enterprises need applications they can scale to thousands of employees and millions of customers. Agents need structured workflows to guide execution. And developers need enterprise-grade software they can trust. All three come from the environment, not the instruction. More code was never the goal. The goal is building enterprise-grade, scalable applications, performant applications, and the harness that closes it.
Ref 1. CodeRabbit, "The State of AI vs. Human Code Generation," December 2025. https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report
Ref 2. Veracode, "2025 GenAI Code Security Report," 2025.
Ref 3. https://smartexe.com/blog/hidden-cost-of-replacing-developers-with-ai