Agent-first IDP: how bigtechs and clouds prepare for agents (Category PlatformEngineering)

Wrote. continuation to an essay on internal platforms. Last time in the textIDP is Dead? No, the GUI monopoly is dying.I’ve explored why the internal platform is losing the monopoly of its GUI and turning into a platform for agents – a layer of capabilities that not only people but also agents turn to. It was about why. The new text is about how.

And in the new article, I specifically did not go into futurology. Playbook is compiled on public engineering materials of companies that have already faced agent consumption of platforms: internal platforms Block, Uber, LinkedIn, Spotify and the revision of public clouds - AWS, Azure, Google Cloud. The logic is simple: when internal bigtech platforms and clouds independently converge around the same engineering solutions, it is no longer fashion, but a new layer of platform architecture.

What this layer is made of: Maturity levels L0 → L4: from GUI-only to controlled agent surface Three layers that are easy to confuse: model gateway, tool gateway, and capability registry Agent identity: the agent is the subject of access on behalf of the user, not a shared service account for AI. Trust levels: read → recommend → act Domain agents where raw access to logs, IAM or SQL is dangerous Evals and telemetry of agent scripts instead of MAU portal A separate threat model: prompt injection, tool poisoning, data exfiltration and more.

At the end, I give you an approximate quarterly plan. 30/60/90 Days and a list of things not worth doing.

The main conclusion is that bigtechs and clouds did not give ready-made instructions that can be copied, but already showed a stable architectural convergence. Internal platforms don’t need to reinvent this path – they need to recognize that their main new user is no longer opening the portal.

Details can be read in my blog

#AI #AI4SDLC #PlatformEngineering #Engineering #Architecture #Management