Engineering Blueprint

Claude Opus + Managing multiple AI coding agents... Playbook

Vince CarolloSoftware Engineer IIIG2April 2026

Multi-agent orchestration system for coordinating specialized AI coding assistants across parallel projects with persistent context and hierarchical task routing.

1 File Included

pasted-prompt.md
16 KB

What problem does this solve?

Managing multiple AI coding agents across parallel projects, coordinating specialized AI reasoning for complex development tasks, and maintaining context across long-running coding sessions.

How does it work?

Three-layer system: (1) Agent of Empires (AoE) - Rust TUI that manages multiple tmux-based OpenCode sessions, allowing parallel agents across different projects with persistent state. (2) OpenCode - terminal AI coding agent routed through LiteLLM proxy to Claude models (Opus for reasoning, Sonnet for execution, Haiku for quick tasks), connected to MCP servers (Atlassian, Amplitude, Metabase) for live data access. (3) Oh My OpenAgent (OMO) - plugin providing 11 specialized agents with hierarchical routing: Sisyphus (primary orchestrator) classifies intent and delegates to specialists (Oracle for hard problems, Explore for search, Librarian for research, Sisyphus Junior for execution, Metis/Prometheus/Momus for planning). Tasks are routed to category-specific models and load domain skills (jira-board (custom), metabase (custom), git-master, frontend-ui-ux, etc). Custom slash commands (/review-pr, /address-pr, /agentnews, /refactor) launch parallel agent swarms for specific workflows.

What's the biggest win?

Ability to run multiple specialized AI agents in parallel across projects while maintaining persistent context, with automatic task routing and delegation reducing manual coordination overhead. Parallel agent swarms for code review (7 agents), planning, and research accelerate complex tasks. Session persistence across terminal disconnects and sleep cycles enables long-running development without losing state.

What should I know technically?

Four MCP servers provide live tool access: Atlassian (Jira/Confluence at g2crowd.atlassian.net, project UW/10204), Amplitude (selective tools: search, query_chart, query_amplitude_data, get_event_properties), Metabase (v0.50 at metrics.g2.com with 19 connected databases including Snowflake, PostgreSQL, BigQuery). Model hierarchy: claude-opus-4-6 (reasoning variants: max/high/xhigh) for orchestration/planning/security, claude-sonnet-4-6 for execution/research, claude-haiku-4-5 for quick tasks. Temperature 0.1, max 200 steps for Abed agent. Custom skills defined in ~/.config/opencode/commands/.

What are the constraints?

Metabase accessible only via custom skill, not direct MCP. Jira-board skill limited: cannot transition statuses, edit fields, link issues, or create Epics. Amplitude tools selectively disabled. Oracle consultant is expensive and used only after 2+ failed attempts. Requires macOS for AoE/tmux setup.