andyrewlee/awesome-agent-orchestrators

What it adds over the summary

The summary already gives the category taxonomy, example tools per category, the supported CLIs, and the mid-2026 watchlist. What it does not carry: the list is organized as a decision tree by intent, there is a companion website with richer per-tool detail, and a handful of the "build your own" primitives carry explicit licences — the detail that matters most to an engineer deciding to adopt versus build on.

How it's organized for selection

The top of the page maps intents to categories rather than presenting a flat taxonomy:

flowchart TD
    Q{What do you want?} --> A[Run several agents, review each diff]
    Q --> B[Keep an agent working while away]
    Q --> C[Split one big job across agents]
    Q --> D[Message an agent, not open a tool]
    Q --> E[Build your own orchestrator]
    A --> A1[Terminal TUI/CLI — live in tmux]
    A --> A2[Desktop & Web — GUI or phone]
    B --> B1[Loop runners: one goal until verified]
    B --> B2[Task runners: pull from tracker/board/schedule]
    C --> C1[Multi-Agent Swarms]
    D --> D1[Personal Assistants — general tasks]
    E --> E1[Infrastructure & Primitives]

A web version at agent-orchestrators.com is stated to hold more detail on each tool than the README.

Selection axes worth knowing before you open it

The entries expose distinctions the summary flattens:

• Coordination cost. Some swarm/loop runners keep no model in the orchestration path so coordination costs zero tokens (bernstein, multi-agent-shogun "no coordination API cost"). Others route every hand-off through a model.
• Sandbox backend for isolation is a recurring axis: local Docker vs cloud via Hetzner, Daytona, Vercel, E2B, Modal, Cloudflare Containers (agentbox, sandbox-agent, background-agents, agenttier on Kubernetes Pods/PVCs).
• Coordination protocols as a distinct layer: Concord MCP, Agent Messaging Protocol (Ed25519-signed envelopes, proven delivery), foremerge (intent/scope declaration over Git), wit (function-level locks via Tree-sitter).

Licences (the ones the page actually states)

Most entries state no licence. Those it does:

• MIT: intentic, no_human, AgentBridge, Mnemoverse (server), iva, Hivekeep.
• Source-available: loki-mode under BUSL-1.1, OtoDock under FSL-1.1.

If you are building on a primitive, assume "unknown, check the repo" for anything not in that list.

Limits

This is a curated link list, not an evaluation: no benchmarks, no "which is best," no maturity signal beyond the watchlist (last-commit dates, checked 2026-07-28) for stalled projects. Licence and language details are inconsistent across entries, so the buildable-primitive shortlist above is only as complete as what the author chose to annotate.