AI Insights: Building an Agentic Developer Platform from Scratch

There’s been a lot of talk about shifting from traditional internal developer platforms (IDPs) to versions driven by AI. But, to date, there are few, if any, examples of this transition being made in practice. It’s something we’ve been working on at Mindera for some time. So, we were really excited when a company asked us to build an AI-driven developer platform for them from scratch.

The platform imperative

Around 90%(1) of organisations have now adopted internal developer platforms (IDPs) to cope with the growing twin challenges of rapid team expansion and complex cloud infrastructure, which have outpaced the ability of traditional, ad-hoc processes to scale. IDPs help to handle this complexity, offering self-service tools that allow developers to deploy code independently without getting bogged down in infrastructure.

However, technology innovation moves increasingly quickly, and the race to optimise AI in infrastructure development is already starting to render IDPs obsolete. That’s because most were built under the assumption they will be handling human-paced development, simple workloads, and a single reviewer per change. Introducing AI breaks all those assumptions at once, completely changing the game.

Built in the right way, integrating AI into a platform can bring significant business benefits. Accelerating infrastructure scalability without growing developer numbers. Commoditising models to further reduce costs. Delivering ownership of infrastructure to strengthen intellectual property and remove licensing agreements. So, were we up for the challenge?

The company that approached us not only wanted their ADP built in six months incorporating AI across all roles (developers, cloud architects and beyond), but also operates in a heavily regulated sector. So this had to be done safely and securely. This demanded a major change in approach to traditional platform development.

Methodology and architectural shifts

One of the changes was transitioning from traditional user stories to spec-driven documentation. This was carried out using the open-source OpenSpec framework, designed to help human developers and AI coding assistants (like GitHub Copilot, Cursor, and Claude Code) plan, build and maintain features without losing context.

Key elements included ‘proposals’, which specify the ‘why’ and the ‘what’, and are written to guide AI agents rather than humans. They explicitly define ‘non-goals’ (what not to build) to eliminate AI hallucinations and log ‘future considerations’ to prevent scope creep – the uncontrolled growth or expansion of a project's requirements or deliverables after a project has started, which can happen regardless of time, budget, or resources.

‘Designs’ (the ‘how’) were also introduced – technical execution documents authored by AI and reviewed by developers. However, even a small proposal could result in a design of 20,000 words or more. So, to reduce this complexity and increase efficiency, large designs were systematically refactored into smaller, manageable specs by splitting up the source proposal, which significantly reduced word counts.

The process meant that user stories and manual Jira ticket writing were eliminated. Tickets were automatically generated, but they were synchronised one-way from GitHub proposals to Jira strictly for stakeholder visibility.

Process and ceremony changes

Key elements of the traditional software development lifecycle (SDLC) were either replaced or refined, including:

• Refinements: In-person refinement meetings were replaced with direct asynchronous reviews on GitHub Pull Requests (PRs), creating an auditable decision trail.
• Dailies and retros: Daily meet-ups shifted from tracking manual code output to reviewing generated designs, outcomes and AI evaluations.

Meanwhile, ‘humans-in-the-loop’ were not required, as AI evaluates its own output. However, humans remained strictly at the gates to approve proposals, designs and deployments to prevent catastrophic failures.

Also a ‘second brain’ was developed, which synthesises project data, transcripts and sprint reviews into a centralised knowledge base. This improves AI context and decision-making over time. It also helped the team split proposals to reduce design word counts without losing key information.

Results and learnings

Making this shift in methodology and process helped us meet the stringent goals set by the client. Deployment times were reduced from hours and days to under 10 minutes. Plus delivery throughput doubled after an initial knowledge-accumulation phase.

The project evolved from an internal developer platform into an agentic developer platform (ADP) featuring a conversational front-door agent that translates developer intent into pre-approved infrastructure patterns. It enables humans and agents to follow the same ‘golden path’ to meet the goal of rapidly deploying infrastructure.

This was achieved on schedule and was, in fact, live in three months, featuring seven resource types, including an app service, SQL and key vault. Plus, all of the releases were checked against company rules, and were signed and traceable, optimising governance and compliance – vital for a highly regulated business.

The platform is the edge

While jumping into the unknown with this project, one thing became clear. AI is an amplifier, not an accelerator. It multiplies weaknesses as well as strengths. If you apply it to something that’s broken, it won’t magically fix it. In fact, it could easily amplify the chaos.

Building an agentic developer platform tailored to an organisation’s specific goals creates a stable, safe and secure environment to deploy and optimise AI in that context, maximising value and impact.

It also means that AI models, which are now universally accessible, no longer deliver a competitive advantage. The platform is now the edge.

Sources:

1. Platform Engineering 2.0: An evolution for the AI era — https://platformengineering.org/blog/introducing-platform-engineering-2-0-an-evolution-for-the-ai-era

Find out more about this groundbreaking project and discover some of the challenges that had to be overcome by watching the full Mindera AI Insights 2026 presentation below.