Agentwise

vibecodingwithphil.github.io/agentwise/

What it is

Multi-Agent Orchestration System for Claude Code. 8 specialist agents (Frontend, Backend, Database, DevOps, Testing, …) work in parallel. Real-time dashboard at http://localhost:3001 (yes, the same port we use). 27+ MCP integrations. Self-improving agents with persistent knowledge. Figma Dev Mode integration. Document upload. Website cloning.

What we kept

• Real-time dashboard pattern. Our /harness page mirrors the dashboard concept: live status, per-agent timing, intervention surface. Different UI, same intent.
• Self-improving agent pattern. Their “agents learn from every task with persistent knowledge base” → our identity/<role>.md cross-mission memory.

What we dropped

• Sandboxed execution to avoid --dangerously-skip-permissions. We solve isolation via branchIsolation and worktrees instead.
• 27+ MCP integrations. Most are tied to vendor APIs we don’t use.
• Figma / website cloning / document upload. Out of scope for an autonomous coding harness.

Differences in philosophy

Agentwise	Papercup
8 fixed specialists	13 customisable role prompts
MCP-heavy	Minimal MCP (just verdict + pi)
Token optimisation via “intelligent context sharing”	Token optimisation via prompt cache (99.94% hit rate)
Workspace sandboxing	Worktree-based git isolation
Polished consumer-facing install (npm create agentwise@latest)	Bash supervisor + per-project config

Honest take

Agentwise is structurally closer to us than any other framework. The fact that we both run on port 3001 was a fun discovery. Where we diverge is opinionation: Agentwise has built-in opinions about what an agent should do (Figma, websites, MCP servers); we built opinions about how the loop should run. Both are valid.

If you want a turnkey multi-agent system that ships features today, Agentwise is the closer-to-product option.

Agentwise homepage.