AI Agents in Software Engineering

AI Agents in Software Engineering

The shift happening in software engineering today, with AI-enabled tools and agents, is fundamentally changing our ways of working. We're moving from AI as a "faster hammer" to AI as a digital colleague, changing how teams build, deliver, and manage software. In just six months, we've watched the industry pivot from retrieval augmented generation (RAG) being  the primary mechanism by which you could do things, to autonomous AI agents becoming the new standard. For software teams, this means rethinking not just how we work, but who we work with.

With insights from Microsoft’s 2025 Work Trend Index and real-world examples, ClearPoint’s CTO Rob Cleghorn and Principal AI Engineer Jonathan Ackerman offer a glimpse into a future where productivity, workforce dynamics, and the very nature of work is being redefined.

Watch our recent episode of ClearPoint Unlock to hear ClearPoint CTO Rob Cleghorn and Principal AI Engineer Jonathan Ackerman discuss how AI agents are reshaping software engineering. They explore the unique opportunities for New Zealand businesses to overcome scale constraints, and the leadership decisions you'll need to make as AI transforms your industry.

Understanding AI agents

An AI agent is fundamentally different from the ChatGPT-style assistants most teams use today. Rather than the typical back-and-forth conversation where you ask for something and get a single response, agents operate autonomously. Give an agent a task, and it will execute multiple steps, make decisions, and use various tools to complete it, much like delegating to a human colleague.

"Instead of that kind of back and forth, you have an agent or AI model set up where you can basically go, ‘Hey, please go and do this for me’," explains Jonathan Ackerman, Principal Engineer at ClearPoint.

The key distinction lies in their toolbox. Agents can be equipped with capabilities to query databases, call APIs, send emails, search the web, or update documents. They decide independently how to use these tools to solve the problem you've presented. It's the difference between asking someone to write you an email versus asking them to handle customer enquiries for the day.


The three phases of AI adoption

Microsoft's recent workforce survey pointed to the emergence of a frontier firm, described as businesses “structured around on-demand intelligence and powered by ‘hybrid’ teams of humans + agents, these companies scale rapidly, operate with agility, and generate value faster.”1 The journey to a frontier firm has three distinct phases, human with assistant, human-agent teams, and human-led, agent-operated.

Phase 1: Human with assistant (Now)

Phase one sees individuals using AI tools to enhance personal productivity. Teams adopt code assistants, content generators, and chatbots, but each person still works independently with their chosen tools. This is where most organisations find themselves today. Developers use AI-powered code completion, professionals leverage ChatGPT for content creation, and productivity gains are real but limited. They can work more efficiently, but they're still just one person. “ It's like giving a faster hammer to some of the engineers. They can work faster and keep going, but it doesn't really scale at the same sort of magnitude as agents,” says Rob Cleghorn, CTO at ClearPoint. 

Phase 2: Human-agent teams (Emerging)

Phase two introduces agents as team members working alongside humans. Agents handle specific responsibilities like quality checking, documentation, or analysis, multiplying team capacity while humans maintain oversight.

Here's where things get interesting. We're already seeing organisations move into this phase, where humans and agents work together as teammates. Consider these real examples:

  • Backlog quality agents that review every ticket against your team's definition of ready, flagging missing information before work begins
  • Documentation agents that automatically update technical docs whenever code changes, ensuring documentation actually reflects reality
  • Retrospective agents that analyse completed tickets, identifying patterns in delays, bugs, or requirement issues to surface insights for team improvement

"We can lift the quality of what we're doing," Ackerman explains, highlighting how agents handle tasks teams often skip due to time constraints.

These agents augment teams rather than replace people, taking on the repetitive, everyday work that engineers may lack the time for or want to off-load.

Phase 3: Human-led, agent-operated (12-18 months)

Phase three shifts humans into leadership roles, leading teams composed primarily of AI agents. People focus on strategy, creativity and coordination while agents execute most operational tasks. The survey suggests we're 12-18 months from humans primarily orchestrating agent teams rather than doing the work directly. Success in this phase will require strong leadership skills, the ability to guide, coordinate, and quality-check agent output becomes paramount.

Practical applications today

Development lifecycle enhancement is where many engineering teams start. Agents can validate requirements completeness, automatically update documentation when code changes, and analyse sprint patterns to identify process improvements. One particularly creative example is a "journey agent" that reads through years of commit messages history to tell the story of a codebase, revealing why decisions were made, when major shifts occurred, and how the system evolved.

“We’ve built agents looking at, are the tickets in your backlog of a suitable quality to enter the pipeline,” says Ackerman. “There’s another one that anytime anyone completes a set of code and merges it back into the codebase, the agent will look at the changes, understand the sum of the code changes, find the documentation for that and automatically update it. The documentation is continually reflecting the actual code.”  

ClearPoint has also developed agents that scan completed tickets at sprint end, analysing all commentary across the lifespan of the tickets. These agents identify issues with requirements, bugs that emerged, or discussions that went in circles. The agent generates summaries highlighting where teams could improve their processes, where technical debt needs attention, or where technology issues slow delivery. The insights feed directly into retrospectives, helping teams understand systemic problems rather than treating each issue as isolated.

However, the applications extend well beyond engineering. Customer service teams use agents to handle routine enquiries (though Klarna's experience with 700 AI agents taught them that customers still want human interaction for complex issues). “We’re taking these agent constructs and applying them to real problems for our customers,” says Ackerman. “Not only in the productivity, automation or software engineering side of things, but actually building them into products for the companies we’re working with.” 

For organisations needing hyper-personalised product recommendations, agents understand individual customer profiles, conduct custom product searches, then evaluate each product against that person's specific needs and wants. They create unique rankings explaining why someone might like a product and what aspects might not suit them, moving beyond generic "customers like you also bought" suggestions to genuine individual matching.

80% of companies use AI, yet few see bottom-line results. Your AI Workforce from ClearPoint breaks through this "AI paradox" by deploying intelligent agents directly into your critical business functions. Gain concentrated value that directly impacts your P&L.


The scale challenge

“What does it mean as a software leader to bring in 10,000 junior devs overnight? How do you keep yourself safe?” says Cleghorn. The productivity potential is enormous, but so are the risks. Without proper governance, you could face quality issues, security vulnerabilities, or simple chaos.

The mythical “10x developer” has become more of a reality today, with agents taking a significant load of a developer and multiplying out - but they do still require guidance and attention.

The challenge multiplies at scale. "AI isn't miraculous, right? It's not suddenly gonna become way smarter than your average developer," Ackerman cautions. "It acts very human-like, if you give it bad inputs or bad instruction, it's gonna probably do a bad job."

Who helps these agents? Who explains what needs doing? Who evaluates their output? Managing this scale requires returning to fundamental engineering principles:

  • Separation of concerns: Agents should have clearly defined responsibilities
  • Infrastructure isolation: Limit blast radius through proper architectural boundaries
  • Clear governance frameworks: Who can create agents? What can they access? How do we review their work?

The challenge isn't just technical, it involves transferring critical knowledge clearly and effectively. Senior engineers have years of experience identifying subtle issues that can cause significant problems later. Teaching agents this kind of insight means clearly documenting best practices and turning implicit, intuitive knowledge into explicit instructions. Effective, practical guidelines help agents recognise and avoid risks before they escalate, making safe scaling achievable.

Opportunities and considerations

New Zealand businesses have an exceptional opportunity to leverage AI agents to overcome traditional scale challenges. Historically resource-constrained, Kiwi firms can now harness scalable intelligence rather than relying solely on human resources. This ability to rapidly innovate and implement bespoke solutions levels the global playing field.

Perhaps most intriguingly, the traditional build versus buy equation is shifting. Organisations are realising the potential to deploy agent teams to create tailored solutions at a fraction of the cost of off-the-shelf SaaS products. Rather than generic, one-size-fits-all tools, companies can build specialised capabilities perfectly aligned with their unique business requirements. This agility allows even smaller firms to innovate quickly and competitively.

Phase zero

So where do organisations start? Begin with low-risk applications that demonstrate value: code formatting, documentation assistance, test data generation. Build experience and confidence while maintaining security standards. The organisations succeeding tomorrow are those learning today – creating frameworks that enable their teams to explore AI's potential within appropriate guardrails.

 “When you've got existing systems and customers, and a lot of responsibility and brand - you have to be a little bit more cautious than just throwing tools and hoping for the best,” says Cleghorn. 

Cleghorn notes that for many organisations, AI agents aren’t entirely new territory, they're effectively another type of cloud service. Most businesses already trust cloud providers to handle sensitive customer data and critical workloads. Starting your AI journey within these established cloud environments allows you to leverage existing security, compliance and governance measures.

Instead of attempting big changes right away, introduce agents in well-understood workflows. Identify tasks your teams already perform in cloud systems, then gradually integrate agents to automate or streamline these activities. This steady, controlled approach builds experience and confidence, ensuring your first AI steps are both safe and productive.

Looking ahead

The inevitable "Phase 4," where agents lead agents, raises important questions about how work might evolve and what roles humans will play. This shift will significantly alter both the skills we value and how teams function day-to-day.

For New Zealand, this represents a unique moment. Our size, traditionally seen as a limitation, becomes an advantage when intelligence can scale instead of people. Our pragmatic, "number 8 wire" mindset aligns perfectly with agent technology, where practical experimentation and innovative problem-solving are essential.

Organisations that proactively support people through this transition, remain adaptable and embrace ongoing learning will thrive. Those who prioritise strong leadership and clearly defined governance from the start will find themselves ready for whatever Phase 4 brings.

Key takeaway

Teams are integrating AI agents into their workflows, transforming how routine tasks are managed and freeing humans to focus on strategy, creativity, and complex problem-solving. The shift towards agents as digital teammates is already reshaping roles, responsibilities, and productivity. Organisations that embrace this transition early will secure competitive advantages, achieving higher-quality outputs and faster results. The question isn't whether to adopt AI agents, but how quickly you can do so whilst maintaining quality and safety.

From improving code quality to keeping documentation up to date, AI agents are already reshaping how software teams work. We’re helping organisations adopt these tools with clarity, care, and measurable results. Discover how ClearPoint can help you take the next step.

 


1https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born

Empower your digital journey