Blog by Paul Listo, Platform Services Lead
The right workflow architecture empowers every engineer to thrive in the agentic era.
The software development lifecycle is a multi-disciplinary business process. I’ve lived in multi-disciplinary processes my whole career: from construction drafting to IT operations, from software development to large transformation programmes. And there’s one thing I’ve learned in every single one of them.
Everyone needs to have their say. And everyone thinks they’re the most important.
That tension is the real story of software delivery. Platform teams, security teams, application teams, business analysts, end users. They all bring legitimate perspectives. They all need to be heard. But the process of hearing everyone, synthesising their inputs and turning that into something a development team can actually build? That’s where things fall apart. Decisions get buried. Assumptions go undocumented. Context evaporates between handoffs. Three sprints later, someone asks “why did we build it this way?” and nobody can answer.
This is the environment into which we’ve introduced agentic AI. And I think most organisations are getting it wrong.
We accelerated the wrong part of the pipeline
The industry collectively decided that the biggest problem in software delivery was that developers weren’t writing code fast enough. So we invested heavily in the coding phase. Autocompletion. Test generation. Boilerplate at scale.
Gartner’s 2025 AI in Software Engineering Survey (1) confirms the pattern. More than half of respondents use AI for writing code, debugging and generating documents. Fewer than half use it upstream in requirements, ideation and design. Fewer still use it downstream in code reviews, deployment and monitoring.
We accelerated the middle of the pipeline and left the ends manual.
There’s a classic tension in multi-disciplinary delivery that anyone who’s lived it will recognise. Don’t shift left enough on quality and there’s too much chaos in delivery. Shift left too aggressively and there’s structure everywhere but no budget to actually build anything. The balance is everything. And AI, as it’s currently deployed, hasn’t touched that balance at all. It’s made developers faster at producing code without improving the quality of what feeds into that code or the governance of what comes out the other side.
McKinsey’s QuantumBlack team (2) observed the same thing from enterprise engagements: individual developers get faster, but the idea-to-live-feature pipeline barely improves. The handoffs between phases are where context goes to die.
AI is a force multiplier for whatever feeds it. Bad inputs, multiplied by AI, produce bad outputs faster. That’s the uncomfortable truth the productivity metrics aren’t capturing.
The human cost of getting this wrong
There’s a harder conversation underneath the productivity one. It’s about what’s happening to the people.
Anthropic published a controlled study earlier this year (3) that should give every engineering leader pause. Developers who used AI coding assistants scored 17% lower on comprehension tests than those who coded by hand. Debugging skills took the biggest hit.
Given how deeply AI is now embedded in the way we build software, that’s a challenging statistic to sit with.
But here’s what gives me genuine hope. The study also found that how developers used AI determined whether they learned or lost skills. Those who asked conceptual questions, who used AI to understand rather than just to produce, scored 65% or higher. Those who delegated blindly scored below 40%.
The problem isn’t AI. It’s how we’re structuring the work around it. And that’s solvable.
What should actually change
If agentic AI is a force multiplier, the question becomes: what should we be multiplying?
I think the answer is collaboration. Specifically, the ability for every discipline involved in software delivery to contribute their perspective in a structured way that feeds directly into what gets built.
Think about what agentic AI can actually do when applied across the full lifecycle rather than just the coding phase. It can help team members communicate their points clearly. It can bring diverse inputs together into coherent, structured documentation. It can surface conflicts between requirements before they become conflicts in code. It can link every software class back to a business requirement so that when someone asks “why did we build it this way?” the answer is right there, traceable from intent to implementation.
The end result is code written from well-considered, balanced and curated documentation. That’s what actually improves productivity. Less politics. Greater collaboration. More time to review and approve what it is we want as an output rather than scrambling to interpret what someone meant three sprints ago.
This is a fundamentally different vision from “AI writes code faster.” It’s AI that makes the entire team more effective by preserving context, enforcing structure and ensuring that every voice in the process is captured and considered.
What this means for engineers
Both Thoughtworks (4) and Anthropic’s 2026 Agentic Coding Trends Report (5) argue that engineering roles are shifting toward agent supervision, system design and output review. I think that’s right, and I think it’s empowering rather than threatening.
Engineers aren’t becoming less important. They’re being asked to do the work that actually matters. Evaluating AI output critically. Delegating with clear intent. Exercising judgement on security, ethics and requirements. These are higher-value skills than writing boilerplate, and they develop stronger engineers.
Senior engineers, in particular, have an opportunity that I find exciting. Their judgement, their architectural thinking, their ability to see around corners is exactly what’s needed to design the governance boundaries within which AI agents operate. They become the people who define the rules of the system. And they become the mentors who help the next generation of engineers develop the judgement to work within it.
Ciklum, referencing an MIT Sloan study (6), noted that early AI adopters saw an initial productivity dip before outpacing non-adopters. That U-shaped curve applies here. The dip is where most organisations currently sit. Getting through it requires investing in people, not just tooling.
The architecture that makes this work
Gartner’s report (1) proposes a calibrated governance model: human-in-the-loop oversight for high criticality, high risk, high complexity work. Human-on-the-loop for lower-stakes tasks. That’s the right framework. The problem is that almost nobody has operationalised it.
The evidence from frontier firms points in a clear direction. McKinsey documented a software modernisation programme (7) where humans were elevated to supervisory roles overseeing structured squads of AI agents. The result was more than 50% reduction in time and effort. The differentiator was the workflow structure, not the AI capability.
Software workflow engines that embed AI agents into structured SDLC processes, with defined approval checkpoints, governance recommendations and human-in-the-loop remediation at critical decision points, are the mechanism that makes all of this work at scale. Seniors define the governance thresholds. Agents execute, recommend and flag. Engineers operate within the structure and develop judgement by experiencing the checkpoints in practice, sprint after sprint.
But the market for purpose-built workflow engines that orchestrate agentic AI across the full software development lifecycle is still nascent. Most organisations are assembling bespoke toolchains, stitching together individual coding assistants, separate governance layers and manual review processes. The fragmentation recreates the same context loss that caused the problem in the first place.
Gartner projects platform engineering teams using AI across every SDLC phase will grow from under 5% to 40% by 2027 (1). That growth demands something most organisations do not yet have.
The question I keep coming back to is this: are we using AI to make developers type faster, or are we using it to make the entire delivery process smarter?
The organisations that answer that question honestly, that invest in the workflow architecture, the governance models and the people to make agentic AI work across the full lifecycle, will be the ones that define the next era of software delivery.
The talent and ambition already exist. What’s missing is the architecture to unlock them.
Paul Listo, Platform Services Lead
Paul streamlines AI, cloud platforms, and IT operations, building intelligent systems that cut noise and accelerate delivery. With 15+ years of experience across Australia and New Zealand, he has founded and scaled start-ups, grown national practices, and led high‑performing teams in regulated industries. He excels at shaping go‑to‑market strategies, creating differentiated offerings, and using strategic partnerships to drive technology adoption and reduce costs. As a trusted advisor, he guides executives through cloud transformation with strong governance, FinOps, and DevSecOps foundations. His recent focus is advancing AI‑enabled operations, including agentic and generative AI, to speed issue resolution and augment workloads. He builds practices from the ground up, combining strategic vision with operational discipline to deliver rapid impact and long‑term capability.
Sources
(1) Gartner, “How to Maximize the Impact of Agentic AI in the SDLC” (January 2026, ID G00843405)
(2) McKinsey / QuantumBlack, “Agentic Workflows for Software Development” (February 2026)
(3) Anthropic, “How AI Assistance Impacts the Formation of Coding Skills” (January 2026)
(4) Thoughtworks, “Preparing Your Team for the Agentic Software Development Life Cycle” (March 2026)
(5) Anthropic, “2026 Agentic Coding Trends Report” (February 2026)
(6) Ciklum, “AI Revolutionizing SDLC in 2026” (January 2026), referencing MIT Sloan research
(7) McKinsey, “Seizing the Agentic AI Advantage” (June 2025)
