Platforms for Humans and Machines: Engineering for the Age of Agents — Juan Herreros Elorza

Juan Herreros Elorza, Team Lead for the Cloud Native Technology team at Banking Circle — a global cross-border payments provider processing over 1 trillion euros per year and serving 700+ regulated financial institutions — argues that the practices that make developer platforms accessible to humans are the exact same ones that enable AI coding agents to thrive. The talk frames agent-readiness not as a new engineering discipline but as a forcing function that makes previously tolerated platform anti-patterns impossible to ignore.

The talk opens with a relatable story of a new developer navigating Banking Circle's internal platform (called Atlas, which encompasses compute, infrastructure, messaging, and observability sub-platforms). The developer must copy pipelines from teammates, discover the right people to talk to on different teams, and wait days for database provisioning — a slow, fragile, human-dependent onboarding process. The central insight: while this workflow is painful for humans, it is essentially impossible for an AI agent. Agents cannot physically walk to another floor, wait a week for a resource, or parse tribal knowledge from informal Slack conversations. Platform anti-patterns that teams tolerate become hard blockers the moment an agent is the primary user.

Five concrete engineering practices are presented as the path forward. First, self-service: every resource provisioning flow must be completable without talking to a specific person. Agents need fully automated, end-to-end paths with no manual approval gates. Second, API-first design: agents excel at calling well-defined, schema-validated APIs. Offering a REST API is the foundation; a CLI wrapper or an MCP server on top are additional layers agents can use. Schema validation ensures agents only submit valid inputs and can iterate in a request-response loop until they succeed. Third, local-first workflows with left-shifted validation: because agents typically run on the developer's machine, feedback must arrive as early as possible. Pushing code to CI only to fail five minutes later on an infrastructure mismatch is unacceptably slow for agentic loops. Everything that can be validated locally should be. Fourth, agent-accessible observability: dashboards and Grafana UIs are useless to agents. Logs, metrics, and traces must be exposed through APIs, CLIs, or MCP servers so agents can close the loop by verifying their own actions programmatically. Fifth, structured documentation: agents perform best when documentation lives close to the code it describes (for small repos) or in a centralized, API-queryable knowledge base (for platform-wide docs). Agent-specific context files — agents.md, CLAUDE.md, copilot_instructions.md — should describe how to build, test, deploy, and verify tasks within a given repository. Skills (codified markdown documents encoding platform conventions) can further guide agent behavior for recurring task types.

The talk also addresses the contribution side: because AI lowers the entry barrier to contributing to platform code, teams should expect more PRs from across the organization. This is a double-edged sword. Platform teams must implement guardrails — policy-as-code for security and compliance requirements — combined with context files that tell agent-assisted contributors how to contribute correctly. The recommended approach is to use hard policies for non-negotiable rules and markdown context files for softer conventions.

On measurement, Herreros Elorza references DORA metrics (change frequency, lead time, mean time to recovery) and reliability indicators such as error rates and traffic behavior, alongside platform-specific metrics like support request volume and developer satisfaction frameworks (SPACE). He urges teams to define what success looks like before making platform changes and to measure whether AI-readiness initiatives actually moved those numbers.

The closing strategic point is notably pragmatic: if a team has previously failed to win organizational buy-in for API-first design or proper documentation, they should use AI agent adoption as the business justification. Executive and individual-contributor attention is at an all-time high on AI productivity. Platform engineers should leverage that moment to finally implement best practices that were always best practices.