Walkthrough: How One Jira Ticket Moves Through an Org of AI Agents

A single ticket, PLAT-412 on rate limiting the public /export API, is followed from a quarterly board goal to merged production code. The author describes a layered design in which agents run inside software-enforced controls, including a policy engine for tool calls, a metered model gateway, and a kill switch that does not depend on agent cooperation. Each organizational seat, from CEO to coder to reviewer, is defined by its inputs, decisions, outputs, escalation rules, and performance metrics, which allows humans and agents to be swapped in the same role. The coder agent works through intake, understanding, clarification, estimation, planning, implementation, self-checking, and handoff, with a rule that every acceptance criterion must map to a concrete test. Hard stop rules such as repeated-error escalation, budget thresholds, and out-of-plan file changes are enforced by the orchestrator rather than by prompts.