Platform engineering & AI agents
Which is local optimisation, literally: one developer, one machine, one afternoon. Wire agents into your SDLC as a capability the whole organisation shares — at the stages where your processes actually run — and they start contributing back into the loop.
The greatest returns on AI investment come not from the tools themselves but from a strategic focus on the underlying organizational system: the quality of the internal platform, the clarity of workflows, and the alignment of teams.
Nathen Harvey, DORA team lead at Google Cloud — The ROI of AI-assisted Software Development
Everyone got faster. Nothing shipped sooner. AI is a tool, and the laws that apply to tools apply to AI — so organisations are walking the same path they walked with cloud, with CI/CD, with Kubernetes.
Enthusiasts prove what is possible — inside their own team, on their own machine. Little interaction between teams, no knowledge sharing, no governance. Wheels get reinvented and the same mistakes get repeated in parallel.
Ideas start to travel, which is real progress. But implementation still happens in the teams: you stop inventing the same wheel and start building it many times over. Then people move, and the versions drift.
When a team finds a new secret sauce, you scale the production of the sauce and put it in the vending machine. Built once, owned by a team, offered to everyone else as a service. Sound familiar? AI is just another tool in the toolbox.
Without a solid foundation built on quality internal platforms and clear workflows, AI merely generates isolated pockets of productivity. That is the DORA finding, and it is what a laptop full of agents produces.
An agent is a new tool and a new capability. Treated as a personal accessory, it optimises one person’s afternoon. Treated as a platform capability, it contributes at every stage your process already has — and feeds what it learns back into the loop.
Notice which stage the industry actually wired up for you.
Seven stages, and the market solved one of them. The rest are where a platform team earns its keep — because nobody else knows your processes, and a generic agent plugged into a generic workflow gives you generic results.
Coding is the part we already made faster, and it is still where nearly all the attention and spend goes. That is local optimisation again — writing code was only ever 20–40% of delivering a feature.
So the bottleneck did not disappear. It shifted left, into requirements and design, and right, into verification, release and operations. Verification especially: when code is cheap to produce, the constraint becomes how fast you can trust it.
Developers move here
Building the understanding of what to build. This is where the leverage is, and it is the side we keep delegating away.
The platform takes the rest
Machines can do this part now — which means the platform has to be built to hold it, not just to hand it over.
Not just build, deploy, run. Now also tickets, documentation, specs and PRDs, designs. If it is context an agent needs, the platform has to reach it — and everything it reaches is untrusted input.
Domain experts prototype and sometimes ship. Product, design, ops and execs are all generating code now, and none of them read your docs.
Agents work on the platform too. That changes platform economics: the opinionated tooling you could never justify hand-building is affordable now, not someday.
Code you can’t verify isn’t value. It’s a liability.
Producing more code than you can check is not delivery, it is inventory with a maintenance bill attached. Bypass the process instead and you ship garbage faster than ever. The answer is neither: tight feedback loops, verification early and automated, and throwaway experiments thrown away quickly. Quality is speed — automating the work does not let anyone skip the thinking.
Platform engineering, extended to cover the two disciplines the agents added: the harness the work runs in, and the loop that starts it.
Human judgement at machine velocity. The trick platform engineering always used — move the knowledge out of people’s heads and into opinionated tools. Rules make a team slower; tools that already implement the rules set it free.
Replacing yourself as the person who prompts the agent, by designing the system that does it instead. Agents embedded in the workflow rather than forced into it locally — triggered by tickets, PRs, branches, scan findings and schedules.
A full development harness, spawned fast and repeatably: versioned skills, MCP configuration, authn/authz, build tools, network restrictions. The coding agent packs its bags and moves out of the editor.
A person’s output was a known quantity. Token spend is not, and agents hold credentials and deploy rights. Limits stop being a judgement call someone makes and become something the platform enforces.
The PRD in OneDrive is not the version your agent sees. Context storage and retrieval is the core place humans and agents now collaborate, which makes an API a requirement rather than a nice-to-have.
Landing zones, infrastructure as code, delivery pipelines and observability on Azure, Google Cloud and UpCloud. AI workloads are not special — the same platforms still mostly work.
Agents are not latency sensitive, are never annoyed by a guardrail, do not need elegant abstractions, and never sleep. Humans notice every extra second, route around friction, and go home. Every capability now has two ends — and most platforms only ever built the human one.
We would rather name these than sell around them. Every one of them is a platform problem before it is anything else.
Cost is variable and unbounded
A person’s output was a known quantity. Token spend is not — one bad loop can make your CFO weep.
Cognitive overload
One human, with agents on a bug fix, a feature and a production outage at once. Nobody context-switches like that.
Review turns into rubber-stamping
Volume outruns attention, and a glance is not verification. Code does not get better just because a human looked at it.
Blast radius
Agents with credentials and deploy rights, working unattended, deploying to production at 3am.
Everything an agent reads can steer it
The platform now reaches tickets, docs and designs. That is untrusted input wired to a machine that can act.
Anarchy by local optimisation
Every team on its own path to production, its own tools, its own stack. Faster individually, unmaintainable together.
Tools don’t sleep. Neither does AI.
So embed the expertise in tools and platform services rather than in meetings, mandates and tribal knowledge — that is how it scales across teams and timezones. And keep the result boring: a page at 2am is not exciting, and neither is a surprise bill from a prototype agent stuck in a loop, burning tokens for no delivered value. Deploying agents should be boring too.
As a certified Google Cloud Partner and UpCloud Partner, we strengthen our Cloud & DevOps delivery with high-performance cloud platforms that enable reliability, scalability and efficiency.
These alliances help us design, build and operate secure, cost-effective cloud environments with measurable impact.
Working together with our clients and partners, we accelerate innovation and create future-ready cloud foundations that support sustainable growth.
Platform strategy, harness and loop engineering, or a sanity check on where your agents are about to hit a wall. Leave your details and we’ll get back to you.