Agent Orchestra

Multiple Claude agents collaborate on your tasks. Each has its own role, context, and expertise. A coordinator manages the interaction across different modes.

Open Dashboard

Modes

Discussion

Round-robin debate. Agents discuss from their perspective, PM summarizes decisions and action items.

Topic → A → B → C → ... → Summary

Pipeline

Sequential handoff. Design, implement, review, test. Loops back on critical issues.

Architect → Developer → Reviewer → Tester

Parallel

Fan-out / fan-in. Agents work on subtasks simultaneously, results merged by coordinator.

Task → [A, B, C] → Merge

Consensus

Structured voting. Each agent votes independently with confidence score. Supermajority decides.

Question → Vote → Tally → Decision

Agents

Architect opus

Developer sonnet

Reviewer opus

PM sonnet

Tester sonnet

DevOps sonnet

Security opus

How It Works

Each agent is an independent ClaudeSDKClient process with its own context window and tool access. Python orchestrator controls turn order, message routing, and stopping criteria. Runs on Claude Max subscription — no per-token API costs.

Roles, prompts, and mode settings are configurable via YAML. Run history and agent sessions persist across restarts.

CLI

orchestra discuss "REST or GraphQL?" --agents architect,developer,reviewer
orchestra pipeline "Add auth" --steps architect:design,developer:implement,reviewer:review
orchestra parallel "Build dashboard" -t "developer:frontend" -t "developer:API"
orchestra consensus "Postgres or Mongo?" --agents architect,developer,devops
orchestra serve   # start web UI

Built with claude-agent-sdk + Python 3.11 GitHub Dashboard