The 3 roles AI agents play in your developer platform
The New Stack Matar Peles
AI agents now show up in developer platforms in three roles: user, workflow step, or managed resource. The tricky part is each role needs different plumbing, and mixing them can get messy fast.
Based on reporting by The New Stack, Matar Peles — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Engineering teams want AI agents inside their developer platforms because speed pressure is relentless. But the same agent can play very different parts, and the article draws a clean line through three of them after hundreds of customer calls.
The first role is the simplest: the agent is a platform consumer. It behaves like another user, pulling context and taking actions. Think Claude Code asking to add an endpoint to a payments service, then reading the service owner, dependencies, and required standards before spinning up a preview environment and running tests. The catch is obvious. If the context is stale or incomplete, the agent can get confidently wrong. The answer, according to the piece, is a governed context layer — ideally a context lake — plus an API- and MCP-first interface and self-service actions the agent can call.
The second role is more industrial. Here, the agent lives inside the platform as part of a workflow, triggered by an event instead of a person. The example is a nightly scan across 40 services that finds vulnerable dependencies, then sends a remediation agent to open a pull request for each owning team. That setup needs orchestration, a registry to fetch the right agent from, a unique identity so actions are logged to the agent itself, and a human-in-the-loop step when the risk is real.
The third role is the most interesting because it treats the agent like infrastructure. The platform provisions it, governs it, and hands it back the way it would a service or database. An engineer can request an on-call triage agent, choose the model, tools, and runtime, and get something back already registered. The article says this is a golden path problem, and that customers are asking for it more than anything else, with agent and skill registry interest up 47% among the organizations they spoke with through early 2026.
The neat part is that these roles can chain together. An agent can be created as a managed resource, later appear as a workflow component, and still read the same governed context as a platform consumer. That’s the real point here: the platform doesn’t need one grand AI mode, it needs a sane way to know which job the agent is doing at any given moment.
My take — AI-written commentary, not fact-checked reporting
This is the part the industry keeps trying to wish away: agents are not a feature, they’re a governance problem wearing a hoodie. The winning platforms will be the boring ones with catalogs, identities, audit trails, and a refusal to let every team improvise its own little robot kingdom. Dry, yes. Also cheaper than cleanup.
Read more about this at: The New Stack