Agentic AI, Coding Agents and the New Dependency Risk

October 13, 2026

Tom Sweet, Head of Technology at SES Secure, explores the rise of Agentic AI and autonomous coding agents, examining the opportunities they present, the new technology dependencies they create, and why operational resilience measures, such as Software and SaaS Escrow remain important.

Over the years, computers and computer programs have proven to be valuable tools for humans. Today, however, the conversation has shifted significantly. The emergence of ‘Agentic AI’ and autonomous coding agents marks a transition from using software as a tool to help create something to the software itself being an active participant in software development and business operations. You must not however, blindly trust in generative output and a human will always be accountable as IBM stated:

“A computer can never be held accountable, therefore a computer must never make a management decision.”
IBM Training Manual, 1979

‘AI Agents’ and machine learning are now capable of generating code which is often promoted to production and can quickly deploy infrastructure. These tools are able to manage workflows and in some cases resolve incidents. ‘AI’ systems are now being trusted with responsibilities that were once designated to experienced technical professionals. From a purely positive perspective, the potential benefits are clear:

  • Faster innovation.
  • Reduced development costs.
  • Improved efficiency.
  • The ability to scale technical capabilities like never before.

However, as organisations work to capitalise on these technologies, a new question is: what happens when artificial intelligence becomes a critical operational dependency?

Agentic AI introduces an entirely new category of operational risk that organisations need to understand before placing their trust in autonomous systems.

The Rise of Agentic AI

Initially, ‘AI’ systems (Large Language Models) operated within narrow boundaries and at first were not even connected to the internet. They could analyse information, generate content, or answer questions in response to user prompts all from a pre-determined data set. This was quickly replaced with newer models that could access the internet and get results from a much wider pool of information and in many cases would ‘hallucinate’ with inaccurate results. Fast forward to today, where Large Language Models are commonplace and are interacted with daily by millions of people around the world. Agentic AI, however, is fundamentally different. Rather than simply responding to instructions, it can independently perform actions and execute tasks to achieve a specific goal without user inputs or prompts.

Coding agents are a good example of the agentic approach. Tools such as GitHub Copilot, Cursor, and Claude Code can:

  • Write code.
  • Test software.
  • Identify vulnerabilities.
  • Deploy changes and.
  • In some cases, coding agents may manage entire development tasks with minimal human intervention.

For organisations, the appeal of coding agents is clear. They can accelerate development cycles and enable smaller teams to deliver output that would previously have required far greater resources.

Remember, as AI capability grows, so too does dependency risk.

Every Technology Revolution Creates New Single Points of Failure

History shows that innovation and dependency go hand in hand. New technologies can deliver significant benefits, but organisations often become reliant upon them before fully understanding the potential consequences of that reliance.

Many businesses experienced this first-hand during the transition from on-premises software to SaaS platforms. The convenience of cloud delivery was undeniable, but organisations went on to discover that they had become heavily reliant upon a single supplier for access to their critical IT assets without a backup or recovery plan in place. Now, businesses are facing a similar shift with AI tools.

Consider a software vendor whose development process becomes heavily reliant on a specific coding agent. Over time, proprietary prompts, AI-generated architecture, automated deployment workflows, and machine-generated code become deeply embedded within the product lifecycle.

The result is a technology ecosystem that may be difficult to understand, maintain or recreate without access to the AI tools and environments used to build it.

The software itself may continue to function perfectly. However, risk factors may emerge when changes, updates, support or future development become impossible because the knowledge required to maintain the system is no longer readily available.

In other words, organisations may find themselves dependent not only on a software supplier but also on the AI infrastructure underpinning that supplier’s ability to operate.

When AI Makes Mistakes at Machine Speed

Another defining characteristic of Agentic AI is speed. Human error has always been a factor in technology failures. The difference is that humans generally make mistakes at a pace that allows intervention. In contrast, autonomous systems can operate at speeds that significantly reduce the window for human oversight.

Recent examples across the technology industry have shown how AI-driven systems can rapidly execute unintended actions. We saw this earlier this year when PocketOS, a company providing software for car‑rental operations, suffered a devastating failure. The disaster was caused by an autonomous AI agent which deleted PocketOS’s entire production database and all associated backups in just nine seconds. 

‘The PocketOS database deletion incident marks a turning point in how digital failures can unfold in the current day. Interestingly, the incident did not involve any gradual, detectable issues. Instead, it unfolded instantly and without any human involvement.’

Why Software Escrow Remains Relevant in an AI-Driven World

At first glance, some may assume that AI reduces the need for Software Escrow. Yet as organisations become increasingly dependent on autonomous systems, continuity and recovery planning become more important than ever.

Software Escrow was originally developed to safeguard organisations against supplier failure by ensuring access to source code, technical materials and other critical assets required to maintain a business-critical application.

While the world of Software Escrow has evolved and changed since its early days, the underlying principle has remained unchanged. What is changing is the scope of assets that may need protection.

As AI becomes embedded within software development lifecycles, business continuity planning may require organisations to consider protection measures for:

  • Source code repositories.
  • Build environments.
  • Infrastructure configurations.
  • Deployment pipelines.
  • Technical documentation.
  • Prompt libraries and workflow configurations.

If a supplier is no longer able to support a solution, these assets can help organisations maintain, recover, or transition critical systems in a controlled and structured way. Software Escrow should therefore be seen not as a simple legal protection measure, but an essential component of modern operational resilience.

SaaS Escrow and the Challenge of Cloud-Native AI Platforms

If an organisation relies on a SaaS solution that incorporates agentic AI capabilities, supplier failure can introduce several layers of risk. Critical AI-driven functionality may become unavailable if:

  • Access to the application is disrupted.
  • Business-critical data can no longer be accessed.
  • Automated workflows and processes stop functioning as intended.

This is why SaaS Escrow is becoming increasingly important. By securing access to the application, data, and supporting assets, organisations are better positioned to maintain continuity and recover quickly if a supplier can no longer deliver its service.

Rather than focusing solely on source code, SaaS Escrow solutions help organisations establish a proven recovery framework that includes software, data and the operational processes necessary to rebuild and restore services in an independent environment.

Innovation and Agentic AI Resilience – Finding the Right Balance

Agentic AI and machine learning technology has already transformed software development, business operations, and digital services in ways that were unimaginable just a few years ago.

The organisations that derive the greatest value from AI will not be those that adopt it without question or those that reject it entirely. Rather, it will be those that harness its potential while maintaining effective controls, robust continuity plans, and strong resilience measures.

Remember, asking difficult questions about dependency risk does not slow innovation, it enables sustainable and considered progress without a cavalier approach to business-critical operations.

Looking Ahead

Agentic AI represents one of the most significant technological shifts since the emergence of cloud computing. It can automate development, act independently and make operational changes in production environments which will impact businesses across every sector.

However, as highlighted throughout this blog, as organisations increasingly rely on AI-powered suppliers, AI-generated software and autonomous systems, continuity planning must also evolve accordingly. Understanding how critical applications can be restored, supported and maintained during periods of disruption will become just as important as the functionality those applications provide.

The future may be built by AI, but resilience will remain a human responsibility. It depends on organisations retaining control of the assets, knowledge, and recovery pathways needed to navigate unexpected disruption.

Tom Sweet has been at SES Secure since 2014. Currently, as the Head of Technology, he leads the Operations team and works closely with partner organisations to develop tailored software escrow and continuity solutions. With a strong understanding of third-party risk, resilience and compliance, Tom helps organisations manage business critical applications and strengthen recovery capabilities.