Context
Multi-agent AI systems are moving into production, but most teams run them blind — no view of which agent is doing what, what tools were called, or where a task is stuck. AgentMesh is a privacy-safe recreation of a real build — the missing layer: mission control for AI teams.
The problem
The failure mode of agentic systems isn't crashes — it's silent wrong answers. Without traces, memory visibility, and a human-approval gate on risky outputs, no serious business can deploy agents on real work.
The approach
I built the full observability surface: a live agent graph (planner decomposing a task into data, analysis, and writer agents, flowing into human review), a real-time trace log with tool calls and guardrail events, per-agent status chips, and a human-in-the-loop approval queue where flagged outputs wait for sign-off.
What the build shows
The console renders a five-agent system mid-task: the data agent querying a warehouse and vector store, a guardrail catching an unverified financial claim, and the escalation landing in the approval table — the exact safety loop production AI needs.
Recreated for confidentiality — a faithful, fully-working mimic of a real production build; client names, branding and data have been fictionalized.