We just shipped four PRs (#122, #124, #125, #129) that fundamentally change how we orchestrate AI agents at StrugCity. The centerpiece: a God View dashboard that gives you real-time visibility and control over autonomous missions.
What Shipped
MissionList now surfaces missions in 'awaiting_approval' state with inline plan summaries. Each mission shows Approve and Reject buttons—no more blind execution. You see what the agent plans to do before it touches production.
MissionControlDispatchForm brings Zod validation, dynamic role and repository dropdowns, and a budget field (max_cost_usd) to prevent runaway spend. The JSON payload preview shows exactly what you're sending to the agent runtime. Sonner toasts confirm success or surface errors inline.
TaskBoard and LiveLog are now wired to server-sent events via useTaskStream. No polling. Green and yellow connection indicators tell you if the stream is live or reconnecting. Every status change, log line, and result arrives in real time.
The most architecturally significant change: spec-first dispatch (SCE-50). The sc-qa agent writes tests first. Those tests become the success criteria for the implementation agent. This inverts the traditional flow—specification precedes code, not the other way around.
Architectural Significance
SSE replaced REST polling. This reduces latency from seconds to milliseconds and eliminates unnecessary database queries. The connection indicator pattern gives operators confidence that they're seeing current state.
Approval workflows introduce a human checkpoint before high-risk operations. Agents plan, humans approve, agents execute. This mirrors how senior engineers delegate: clear intent, explicit approval, autonomous execution.
Spec-first dispatch formalizes TDD for agent teams. The QA agent becomes a specification writer. The implementation agent has unambiguous success criteria from the start. This reduces rework and makes failure modes explicit.
What's Next
We're extending the approval workflow to support conditional approvals and time-based expiration. Budget enforcement will move from UI validation to runtime checks. And we're exploring agent-to-agent handoffs—one agent requests another's output as input to its own plan.
Follow along at strugcity.com/progress. #StrugCity #AIAgents #BuildingInPublic