IDPs Are Dead: Platform Engineering in the Age of AI Agents - Tiago Reichert & Lucas Duarte, AWS

Summary

The speakers argue that internal developer platforms (IDPs) are not dead. Instead, developers are increasingly accessing platform capabilities through AI coding assistants and agents rather than relying primarily on portals and self-service interfaces. The platform remains important as the place to define approved services, templates, tools, and operational boundaries.

Key concepts and analysis

• Assistants versus agents: An assistant responds to requests, while an agent is given a goal and repeatedly reasons, observes, selects tools, and acts toward it. Because an agent’s exact actions can be nondeterministic, it needs stronger governance than a conventional UI workflow.
• Personal versus production agents: Personal agents act on behalf of an individual. Production agents are treated more like platform-owned services: they need durable ownership, identification, audit trails, and reassignment if their creator leaves.
• Agents as platform workloads: Teams can extend an existing platform, such as Backstage, with templates for creating and managing agents. The platform catalog can constrain which MCP integrations, skills, and other tools agents may use.
• Knowledge and context: The speakers distinguish relatively stable documentation, such as runbooks, which can be retrieved with RAG, from live or situational information, such as current cluster logs, which can be fetched through MCP integrations. They also emphasize capturing tacit organizational knowledge that is not already documented.
• Ontologies and knowledge graphs: Structuring relationships among services, owners, deployment guidance, troubleshooting steps, and approvals can help agents retrieve more relevant context. The speakers say this can reduce token use and repeated reasoning, improve precision, and support governance.
• Identity, authorization, and auditing: Agents should use their own identities rather than a developer’s credentials. Access decisions should also account for the requesting user’s permissions, so an agent cannot use its broader capabilities to bypass the user’s authorization.
• Governance and autonomy: Agent workflows can range from read-only assistance to proposing changes for human approval or acting autonomously. The speakers recommend static guardrails—for example, Open Policy Agent (OPA) policies—and human review for risky changes. They also describe using an LLM as a risk judge to flag cases not covered by predefined rules.

The presenters cite a Spider 2.0 benchmark example in which a model’s performance fell from around 90% on a standard benchmark to 58–70% on a corporate schema when it lacked organizational context. They use this to argue that missing context, rather than model capability alone, is a major obstacle.

Demo and implementation walkthrough

The recorded demos show a Backstage-based catalog for agent templates, integrations, skills, and knowledge sources. In one example, an agent scales a checkout service to eight replicas for a user named Sam. A request to view pods in the restricted kube-system namespace is denied. The speakers present these examples as demonstrations of user-aware authorization, policy enforcement, and auditability.

The closing architecture includes:

1. Agent definitions generated from catalog templates as YAML and stored in GitHub.
2. Argo CD deploying agents from the repository rather than allowing direct cluster deployment.
3. Amazon Bedrock as the model provider in the example.
4. MCP integrations exposed through an agent gateway.
5. OPA with Rego policies for access control, alongside monitoring and audit records.

Q&A and resources

In response to a question about outdated knowledge graphs, the speakers recommend being deliberate about what knowledge agents can use. One suggested workflow is to identify useful patterns from a knowledge graph and turn them into maintained agent skills. They note that instructions and skills may need to be updated as processes change.

The talk includes practical demos and an implementation overview, but no product review. The presenters say a QR code provides access to the platform example, repository, and a book with further details.

Main speakers and sources

• Lucas Duarte — Lead agent transformation specialist at AWS.
• Tiago Reichert — Container specialist supporting startups in Latin America.
• Sources discussed: A Cloudflare study on agent-related internet traffic and the Spider 2.0 benchmark, as cited by the speakers.