AgentSys is a modular runtime and orchestration system for AI agents — a marketplace and installer that ties together 24 plugins, 49 agents, and 44 skills into gated development pipelines — built for developers and teams working in Claude Code, Codex CLI, OpenCode, Cursor, or Kiro who want the work around code generation automated, not only the code generation itself.

What it is

AgentSys is a modular runtime and orchestration system for AI agents, written in JavaScript and released under the MIT licence. It is the marketplace and installer for 24 plugins, each of which lives in its own standalone repository under the agent-sh organisation; installing AgentSys fetches those plugins automatically from their repos. Across the ecosystem the project counts 49 agents — 39 file-based plus 10 role-based specialists in audit-project — and 44 skills, with roughly 30,000 lines of library code and 3,518 tests behind it. It runs on five platforms: Claude Code, Codex CLI, OpenCode, Cursor, and Kiro.

The problem it targets is the work surrounding code generation rather than the generation itself. The project's own framing is that AI models can already write code, and the hard part is everything else: task selection, branch management, code review, artifact cleanup, CI, PR comments, and deployment. AgentSys replaces hand-run, per-session agent prompting and manual phase tracking with structured pipelines that enforce phase gates, so an agent cannot skip a step, and with persistent state that survives session boundaries, so interrupted work resumes instead of restarting.

Key capabilities

• Marketplace and installer for 24 plugins, each distributed as a standalone repository under the agent-sh org and fetched automatically at install time.
• 49 agents, each with a single responsibility, a specific model assignment, and defined inputs and outputs — 39 file-based agents plus 10 role-based specialists in audit-project.
• 44 skills that compose into structured pipelines with enforced phase gates between steps.
• Persistent state across session boundaries, so work survives interruptions.
• Certainty-graded findings: HIGH means definitely a problem and is safe to auto-fix, MEDIUM means probably a problem and needs context, LOW means possibly a problem and needs human judgment.
• Hybrid execution: regex, AST analysis, and static analysis do detection, while LLM calls are reserved for synthesis, planning, and review — reported as 77% fewer tokens for /drift-detect versus multi-agent approaches.
• Commands such as /drift-detect, /can-i-help, and /onboard, plus CLI entry points agentsys install, agentsys --tool codex, and agentsys --tool opencode.

Who uses it and how

• Developers and teams in Claude Code, Codex CLI, OpenCode, Cursor, or Kiro who want task selection, branch management, code review, artifact cleanup, CI, PR comments, and deployment handled by agents rather than manually.
• Teams standardising on more than one assistant, installing the same plugin set per tool through agentsys --tool codex and agentsys --tool opencode.
• Windows and Linux users running the suite or its CI, since the test suite now runs on both platforms.
• Cost-sensitive teams choosing a model tier: the published benchmark on glide-mq puts Sonnet with AgentSys at $0.66 per task against $1.10 for Opus without it.
• Maintainers of larger codebases who want the audit-project specialists, a set exercised against more than 1,000 repositories during development.

Getting started

Install through the marketplace or the npm installer; agentsys install then fetches the plugins automatically from their repositories. On Codex CLI or OpenCode the install path is agentsys --tool codex or agentsys --tool opencode.

How it compares

No competing orchestration runtime or list of paid products it replaces is given in the facts, so AgentSys stands alone in this registry. The adjacent tools named alongside it are complementary rather than competing: agnix is a CLI and LSP linter that validates agent configuration against 423 rules, and tiyuvta inference is hosted, OpenAI- and Anthropic-compatible LLM inference aimed at agents running continuously.

When to use it — and when not to

AgentSys fits teams already committed to one of its five supported platforms and willing to pay for LLM calls per task, since the published benchmark puts a single run between $0.66 and $1.95 depending on model. It is a poor fit for anyone outside those platforms. The install path has also been unstable: version 6.0.2 had to delete two adapter install.sh scripts that destroyed a working install while reporting success, and had to fix Windows execution by resolving the Claude Code executable with where.exe instead of assuming claude.cmd. Teams that cannot absorb that kind of churn in an installer should treat the release line as young.