Infobip publishes the four-phase workflow its AI team uses to operate coding agents, front-loading human review

arXiv·medium signalXBlueskyLinkedInCopy link

This is a practitioner report rather than a benchmark: Infobip's AI research team structures agent-assisted development into four phases where human effort is heaviest early and delegation increases as artifacts mature, with four context-management strategies applied at each phase against known failure modes. Their observation is that upstream errors in research and planning compound through later phases, while correcting already-generated code introduces bloat and fragility, which is the argument for reviewing the plan rather than the diff. They name two open problems honestly: there are no metrics for workflow effectiveness, and formalized context-management components do not map to the workflow-level patterns practitioners actually need.