Eduardo Simón Picón and Javier Casal Ruiz explain why Flywire built agent infrastructure around its existing engineering platform. Orbit expresses workflows as reviewed YAML, keeping predictable procedures in code and using a model only where reasoning is needed. A code-review example combines deterministic context collection with a read-only model step, structured findings and a final posting step.
Eduardo Simón Picón and Javier Casal Ruiz describe separate orchestration, execution and session-state components. Timeouts, rate limits and task locks constrain resource use and duplicate execution. Each model step declares its network, tools, skills and permitted MCP operations, while merge requests retain a human approval boundary. Running internally is presented as part of their governance design, not independent proof of secure data handling.
Eduardo Simón Picón and Javier Casal Ruiz emphasize reusable skills and strict JSON output schemas. They test individual model steps against recorded historical context rather than relying only on noisy end-to-end benchmarks. Production tracing feeds versioned golden datasets, but they acknowledge that initial acceptance still often relies on a few practical cases and that online evaluations are not yet in place.
Eduardo Simón Picón and Javier Casal Ruiz report applications including code review, dependency fixes, translations, incident investigation and repository maintenance. Their execution counts and success rate are internal operational claims, not independently validated measures of code quality, security or downstream correctness. Choosing models per task and separating engineering, corporate and payment-production data remain ongoing responsibilities.
Eduardo Simón Picón and Javier Casal Ruiz identify full sandboxing, resumable sessions, richer evaluations and model- and harness-independent orchestration as roadmap work. They explicitly state that the current runtime is not fully sandboxed. Their practical recommendation is to start with a bounded repetitive task, keep deterministic work in code and retain ownership, limits and human review as automation scales.
Watch on YouTube
