Agent Orchestration
If you're on-call, one Claude isn't enough.
You're on-call at AWS. Four tickets are open, each one in a different console — Bedrock, S3, CloudWatch, Lambda — and every console is its own MFE in its own repo (even services within Bedrock are fragmented). To investigate in parallel with a single-session CLI, you'd be shuffling between terminals, juggling context, and taking scratch notes every time a subagent comes back.
These four prototypes ask: what would it feel like if the harness knew you were orchestrating? Each one dispatches the same four AWS triage agents and lets you jump between them — with recap, current task, next task, and a live log — in a different shape of surface.
Prototypes
Tabbed Agents
Four agents tabbed on the left, each with a full claude-style chat surface on the right. Tabs show status (working / needs you / idle) and a one-line ask preview. The right side shows recap, current task, next task, and an inline log of what the agent is actually doing.
CLI Subagents
Claude Code’s current subagent output, but interactive. Each subagent row shows its active tool call underneath, and you can "dig in" to a subagent to see the full recap / task / log and talk to it directly.
Popup Dashboard
Agents "come back" to the dashboard as avatars when they need a decision. Click one and a chat box opens with their question. Actively-working agents stay pinned at the bottom as a progress strip.
Repo Map
Visualize repos and folders as a radial tree and place each agent as an avatar on the folder it’s touching — kind of like a video-game overworld. Idle agents queue in a bar along the bottom; click to prompt them back into action.
Shared scenario
Every prototype consumes the same mock roster so you can compare surfaces apples-to-apples. The roster lives in lib/orchestration-scenarios.ts.
Bedrock
p99 InvokeModel latency up 4× after deploy
S3
ListBuckets 500s for accounts with >1000 buckets
CloudWatch
bedrock-5xx alarm flapping every 90s
Lambda
Customer: deploy rollback UI "hung"