Agents in the Wild · Session 4 · In-person
Orchestration Frameworks
March 13, 2026 · The Kannas Hotel, Chiang Mai
One agent is useful.
Who coordinates the team?
The Orchestration Problem
- Managing one agent is too slow — multi-agent seems to be the evolution
- Multi-agent = new questions: coordination (how?), visibility, structure (what?)
- Everyone is solving this differently — no standard yet
- The spectrum: from terminal panes to full company simulation
The Spectrum
lightweight / terminalheavyweight / platform
tmux-ide
YAML + panes
dmux
worktrees + agents
Paperclip
org charts + budgets
Codex
cloud agent platform
Symphony
multi-agent framework
Different tools for different levels of autonomy.
The orchestrators
tmux-ide
terminal orchestration. What: Multi-agent teams in tmux panes. Lead agent spawns teammates on-the-fly. Shared task list for coordination. How: Define team layout in ide.yml — One command launches everything. Auto-detects stacks: Next.js, Vite, Python, Go. Trade-off: Lightweight. Terminal-native. No web UI. No budgets. No governance. Best for: devs who live in the terminal. tmux-ide · github.com/wavyrai/tmux-ide
dmux
parallel agents. What: Full TUI for parallel agents. Each agent gets its own git worktree + tmux pane. 11 supported agents. By Justin Schroeder (FormKit). 1k stars. v5.6. How: Press n → type prompt → pick agents. Worktree created, branch created, agent launched. Press m to merge back. Built-in file browser, diff preview, conflict resolution. Trade-off: Most polished TUI in this space. Git-native isolation. Smart merge flow. But: dev-focused. No business logic layer. Best for: shipping code with multiple agents. dmux · github.com/standardagents/dmux
Paperclip
the full platform. What: "Hire AI employees." Org charts. Budgets. Goals. Governance. Web dashboard at localhost:3100. Architecture: Node.js + React + Postgres — 7 agent adapters: Claude, Codex, Gemini, Cursor, OpenCode, Pi, OpenClaw. MIT licensed. Self-hosted. 21k stars in 11 days. Trade-off: Full control plane. Approval gates. Cost tracking. But: heavier setup. Agents inherit server env. Best for: teams wanting structure. Paperclip · github.com/paperclipai/paperclip
Live Demo
Paperclip
Running in Docker on this machine right now. Node.js server + embedded Postgres + React UI. Agent spawns CLI processes (Claude Code, Codex) inside the container. Everything sandboxed — nothing touches the host. localhost:3100
The orchestrators, continued
OpenAI Codex
cloud agent platform. What: Cloud-hosted coding agents. Each task runs in its own isolated sandbox. Multiple agents in parallel. 1-30 min per task. No internet access during execution. How: "Command center for agentic coding." Reads/edits files, runs tests, linters, type checkers. Proposes PRs when done. Web app + CLI + VS Code — same backend. Trade-off: Fully managed. No infra to run. But: closed platform. OpenAI models only. Your code goes to their cloud. Best for: teams wanting zero setup. OpenAI Codex
Symphony
OpenAI's orchestration spec. What: Orchestration system that turns project boards into autonomous implementation runs. Monitors Linear/GitHub, spawns agents per task. Key Idea: "Manage work, not agents." Agents deliver proof of work: CI status, code review, complexity analysis. Full lifecycle: task → agent → PR → landed. Trade-off: It's a spec, not a product. Reference implementation in Elixir. Build-your-own or follow the pattern. Best for: teams ready for "harness engineering". github.com/openai/symphony
Custom builds
People Are Building Their Own
"I moved from TUIs/IDEs to my own agent orchestrator in 3 months." — @omarsar0: custom dashboard with sessions, tasks, skills, automations, workrooms. Auto-ingests arxiv papers. Maps X trends to research. Agents run 24/7. The message: the tools don't do everything you need. Build the missing pieces.
— @omarsar0
in an orchestrator
What to Look For
- Isolation — can agents trash each other's work?
- Coordination — shared tasks? message passing? or just hope?
- Approval gates — does a human review before the agent acts?
- Cost visibility — do you know what each agent is spending?
- Security — what can a compromised agent access?
Open floor — shout it out.
What matters most to you?
- ›What would you look for in an orchestrator?
- ›What's missing from what you've seen today?
Where This Goes
- Today: developer tools — agents help you code
- Next: business tools — agents run workflows
- Eventually: agents coordinating agents — no human in the loop
- The orchestrator is the operating system for this future.
Discussion · Demos · Q&A
Which layer are you building at?
Follow the lab.
See you next Friday.