Stop building new AI agent orchestrators – there are 636

Open-source developers have created hundreds of redundant frameworks to coordinate coding agents, according to metrics compiled by github.com's report. The repository tracks 636 open-source and source-available tools designed to run, supervise, or coordinate artificial intelligence agents across terminal, desktop, and container environments. The catalog originated from andyrewlee's awesome-agent-orchestrators list under a CC0-1.0 license, adding 275 new repositories and 15 adjacent frameworks to monitor how the developer ecosystem manages concurrent autonomous agents.

This analysis examines the operational patterns, feature distribution, language choices, and maintenance risks across the tracked repositories. It is intended for software engineers, systems architects, and engineering managers evaluating whether to adopt existing agent tooling or build custom runners. The collected data shows that the core technical requirements—session isolation, multi-agent multiplexing, and inter-agent communication—are already implemented across dozens of active projects, making community contributions and stability far more urgent than writing another orchestration framework from scratch.

01Ecosystem scale and repository activity

The tracked catalog represents an aggregate footprint of 3,684,348 GitHub stars distributed across 636 projects as of November 15, 2026. The median repository in the dataset has seven contributors and recorded its last commit just three days prior to collection. Activity classifications define active projects by a commit on the default branch within 30 days, recent projects within 90 days, slow projects within one year, and stale projects beyond that window.

Under these standards, 449 repositories maintain active status, while 69 rank as recent, 108 are classified as slow, three are stale, four are archived, and three are completely unavailable. The overall catalog divides into specific operational categories, led by desktop and web tools with 188 projects and 568,306 stars, multi-agent swarms with 114 projects and 632,599 stars, and infrastructure systems with 82 projects holding 255,949 stars. Terminal interfaces account for 61 tools with 130,390 stars, while task runners hold 58 projects and 244,774 stars.

Personal assistants comprise 57 projects with 1,110,011 stars, loop runners track 31 projects with 63,657 stars, adjacent frameworks encompass 28 repositories with 630,893 stars, and 17 resting projects collect 47,769 stars. Across these nine categories, active or recent maintenance rates remain exceptionally high, exceeding 85 percent in terminal, desktop, infrastructure, assistant, and adjacent framework groups.

02Mechanics of concurrent agent isolation

Most surveyed orchestrators implement nearly identical mechanical layers to execute tasks. Their core operational loops focus on running several agent sessions side by side, separating file access through distinct environments, and routing programmatic steps between worker agents. Repository feature analysis reveals that 45 percent of tracked tools rely on Git worktrees for workspace isolation, while 41 percent employ containers, 38 percent utilize sandboxes or virtual machines, and 21 percent use tmux sessions.

Agent engine support across project documentation shows heavy concentration around mainstream coding assistants. Claude Code appears in 78 percent of project descriptions, followed by Codex at 74 percent, OpenCode at 49 percent, Cursor at 39 percent, and Copilot at 25 percent. Gemini CLI accounts for 23 percent of mentions, while Qwen Code, Kiro, Aider, and Cline each appear in seven percent, Goose in five percent, and Amp in three percent.

User interfaces and protocols reflect standard developer deployment targets. Desktop applications lead interface formats with 35 percent mention share, followed by web user interfaces at 29 percent, terminal text user interfaces at 25 percent, mobile layouts at 14 percent, and integrated development environment extensions at two percent. In communication standards, Model Context Protocol adoption reaches 55 percent, Agent Client Protocol sits at 15 percent, and Agent-to-Agent protocols register five percent, alongside scheduled run support at 21 percent.

03Languages and leading project stars

TypeScript serves as the dominant implementation language across the tracked orchestrators, powering 295 repositories or 47 percent of the entire catalog. Python ranks second with 108 projects accounting for 17 percent share, while Rust powers 74 tools at 12 percent. Go represents 60 projects at 10 percent, JavaScript covers 34 projects at five percent, Shell scripts power 21 projects at three percent, Swift holds 15 projects at two percent, Elixir claims five projects, Zig has four, and C-sharp accounts for three.

Project popularity metrics place openclaw at the top of the overall ecosystem with 394,509 stars, 376 contributors, and an active status within the assistants category. The hermes-agent assistant ranks second with 263,549 stars and 399 contributors, followed by the Orca desktop platform with 125,454 stars and 464 contributors. Within swarm orchestration, paperclip holds fourth place with 114,052 stars and 243 contributors, while OpenHands leads task runners with 93,506 stars and 456 contributors.

Other high-visibility repositories include DeerFlow with 85,363 stars and 454 contributors, ruflo with 81,777 stars, MetaGPT with 71,280 stars, OmO with 70,917 stars, AutoGen with 61,844 stars, t3code with 61,078 stars, crewAI with 60,940 stars, multica with 56,175 stars, herdr with 51,090 stars, and nanobot with 50,099 stars. High-activity commit logs on November 15, 2026 recorded 48,762 ninety-day commits for openclaw, 12,798 for Orca, 8,309 for Hive, and 6,234 for Taskuary.

04Terminal runtimes and multiplexer tooling

Terminal-based orchestrators represent 61 projects, where developers interact directly through command-line dashboards and multiplexer panes. Within this category, herdr leads with 51,090 stars and 87 contributors as an Apache-2.0 licensed background runtime written 94 percent in Rust, maintaining version 0.9.5 to keep agent terminals persistent across operating system reboots. The cmux project provides a Ghostty-based macOS terminal with vertical tabs, logging 28,790 stars, 250 contributors, and release cmux-app-ffi-e238ea5486824ff4fc766954d4bc91d37b90af0c on November 15, 2026.

Go-based terminal managers provide lightweight background process isolation. The claude-squad utility holds 8,811 stars under the AGPL-3.0 license, using Go to detach Claude Code and Codex sessions inside separate Git worktrees under version 1.0.20. The newer tuios multiplexer tracks 6,532 stars and 5,661 ninety-day commits, delivering version 0.13.1 with built-in Model Context Protocol and Agent Client Protocol support across interactive terminal panes.

Additional terminal environments cater to remote monitoring and hardware-specific workflows. The agent-of-empires project pairs an AGPL-licensed Rust terminal dashboard with mobile web views, registering 4,472 stars and version 1.24.0. The abtop tool monitors session tokens and port bindings across 3,857 stars at version 0.5.5, while Toad uses Python to coordinate agents over the Agent Client Protocol with 3,503 stars at version 0.6.20, and Gentle Shell delivers version 4.8.0 for Raspberry Pi workspaces across 2,657 stars.

05Maintainer distribution and single-author vulnerability

Despite strong aggregate contributor numbers, a significant portion of active agent orchestration software depends entirely on solitary developers. The dataset tracks a dedicated single-maintainer risk metric identifying projects where one individual author contributed 100 percent of all human commits across repositories with at least 100 non-archived commits.

Ten prominent active or recent projects exhibit total reliance on a single author. The Mercury CLI repository records 13,693 all-time commits under active maintenance by one contributor, while pohunek maintains 4,463 commits and tmux-ide tracks 4,278 commits under identical single-person ownership. The amux project records 1,126 commits and Calyx logs 967 commits, each managed entirely by one developer without secondary committers.

Similar single-maintainer exposure appears in pando with 528 commits, Craig with 307 commits, OpenCode Mission Control with 154 commits, taskpods with 129 commits, and openkanban with 118 commits. This concentration leaves critical isolation and workflow scripts vulnerable to abandonment, indicating that engineering teams adopting agent infrastructure must weigh contributor depth, active test suites, and documentation maturity against raw GitHub star counts.