AgentSys
A gated, resumable delivery system for turning development tasks into reviewed and merged changes.
Per-dimension scores and reasoning
Evidence shows basic security measures: SECURITY.md mentions command injection prevention, path traversal prevention, and input validation, but no implementation details. User confirmation: /next-task includes user approval step, /ship has --dry-run, but other commands like /deslop apply may modify files directly. Data flow transparency limited, no detailed explanation of data collection and transmission. Sensitive data handling: SECURITY.md reminds users not to commit secrets, but no system-level handling. Dependency security: 6.0.1 removed vulnerable js-yaml, but no full audit. External effects: /ship and /next-task commit, push, merge PRs, affecting external systems, but provide --dry-run. Rollback: no explicit rollback mechanism, but /next-task has --abort. Source attribution: plugins from agent-sh org, but publisher unverified.
Self-consistency: README claims 24 plugins, 49 agents, 44 skills, consistent with description, but not verified against actual files. Dependency availability: depends on external tools like agnix CLI, gh CLI, but no handling for missing tools. Failure messages: no specific error handling docs, but tests cover some edge cases.
Audience clear as developers, scenarios cover code development, review, release. Capability boundaries: each agent has single responsibility, but no explicit limits. Trigger precision: commands have clear parameters, but no trigger condition details. Environment fit: supports multiple platforms, but no detailed configuration guide.
Information architecture clear, README has TOC and command list. Install notes provide npm installation. Naming stable, command and agent names consistent. Examples and FAQ: README provides command examples, but no FAQ. Known limitations: SECURITY.md mentions no security guarantee, but no specific limitations. License MIT, version 6.0.1, CHANGELOG exists but not provided. Maintenance responsibility: author Avi Fenesh, but no maintenance plan.
Output usability: command outputs have clear formats, e.g., /deslop report. Marginal value: provides various automation commands, saving time. Cost benefit: README provides benchmarks showing Sonnet + agentsys cheaper and comparable quality to raw Opus.
Claim traceability: README claims 3518 tests, but no test results provided. Cross-source corroboration: no external verification. Fact/inference separation: README distinguishes benchmarks and inferences, but not explicitly labeled.
- Publisher identity unverified, use with caution.
- Agents can perform modifying operations (e.g., /ship, /deslop apply), always review changes before running.
- Depends on external tools (e.g., agnix CLI), ensure they are installed.
- Security measures are limited, no implementation details provided.
What does this agent do, and when should you use it?
AgentSys is a modular runtime for orchestrating software-development agents and a marketplace/installer for plugins under agent-sh. It combines 24 plugins, 49 agents, and 44 skills behind slash commands for task discovery, implementation, review, documentation sync, release, and merge workflows. Its documented pipelines use phase gates and persistent state, and `/next-task` creates isolated worktrees for task execution. The system is documented for Claude Code, Codex CLI, OpenCode, Cursor, and Kiro, with an npm-based installer. It fits teams that want repeatable engineering delivery workflows, provided they are comfortable with CLI, Git, and platform-specific workflow dependencies.
With /next-task, AgentSys selects work from GitHub Issues, GitHub Projects, GitLab, or a local file; creates a worktree; explores the repository; prepares a plan for user approval; implements it; runs deslop and test-coverage checks; iterates through review; validates delivery; syncs documentation; and invokes /ship for PR creation, CI monitoring, and merge. /prepare-delivery runs pre-review gates, conditional config linting, a review loop, delivery validation, and docs sync without pushing or opening a PR, while /gate-and-ship composes it with /ship. /repo-intel uses agent-analyzer to collect Git-history intelligence, AST symbols, and project metadata into platform-state caches; /drift-detect gathers issues, documentation, and code exports with JavaScript collectors before semantic analysis. Other documented commands include /audit-project, /perf, /sync-docs, /agnix, /onboard, /consult, and /release.
- A GitHub-based engineering team that wants a selected issue to move through planning approval, implementation, review, and merge in one workflow.
- A developer with finished local changes who wants
/prepare-deliveryto run cleanup, coverage checks, multi-role review, validation, and documentation sync before deciding whether to open a PR. - A repository maintainer who needs
/shipto detect CI and deployment tooling, push a branch, create a PR, follow CI, address review comments, and merge. - A contributor joining a JavaScript, Rust, Go, Python, C/C++, Java, or Deno project who needs
/onboardto summarize its stack, key files, and conventions. - A team maintaining skills, agent prompts, hooks, and tool configuration that wants
/agnixvalidation for structure, security, consistency, and cross-platform configuration rules. - A project owner who suspects issues, plans, or documentation examples have diverged from implementation and needs
/drift-detector/sync-docs.
What are this agent's strengths and limitations?
- Composes task discovery, isolated worktrees, plan approval, implementation, review, validation, documentation sync, and PR merge into standalone or chained commands.
- Uses specialized responsibilities, phase gates, and persistent state for long-running delivery workflows that can be resumed.
/repo-intelcentralizes cached Git history, AST symbols, and project metadata for reuse by multiple plugins./agnixdocuments 423 validation rules and SARIF output, enabling agent-configuration checks in GitHub Code Scanning.- Provides documented installation paths for Claude Code, Codex CLI, OpenCode, Cursor, and Kiro.
- It requires Git and Node.js 18+, while GitHub and GitLab delivery workflows also depend on authenticated
ghorglabclients. /shipis designed to commit, push, create PRs, monitor CI, address comments, merge, and clean up; that scope may not fit teams that do not want automated remote repository changes./prepare-deliveryis the documented non-shipping alternative.- Some functions have added tooling dependencies:
/agnixneeds the agnix CLI, and/repo-intelrelies on agent-analyzer installed through npm. - The reported benchmarks are tied to named tasks, repository, and model configurations, so they do not establish equivalent cost or quality outcomes for every project or model.
How do you install or deploy this agent?
Git and Node.js 18+ are required. Install the cross-platform npm installer with npm install -g agentsys && agentsys. For Codex CLI, run agentsys --tool codex; for Claude Code, run agentsys --tool claude. Claude Code can also use /plugin marketplace add agent-sh/agentsys, followed by a plugin install such as /plugin install next-task@agentsys. GitHub workflows require an authenticated gh; GitLab workflows require an authenticated glab. /agnix additionally requires the separately installed agnix CLI.
How do you use this agent?
After installation, run /onboard for an overview of the current repository. Run /next-task for the full task-to-production workflow; it requests plan approval before implementation. Run /prepare-delivery for quality gates without creating a PR, then /ship for commit, PR, CI, and merge handling, or use /gate-and-ship to chain both. Resume an interrupted task workflow with /next-task --resume. Initialize repository intelligence with /repo-intel init, then query it with /repo-intel query hotspots.
How does this agent compare with similar options?
The README compares its structured pipeline with direct model use without AgentSys. For the documented documentation-improvement task, it reports Sonnet + AgentSys as comparable in result quality to Opus + AgentSys, while raw Opus produced more generic, less project-specific output. /drift-detect is also explicitly positioned against multi-agent collection: it uses JavaScript collectors plus one semantic analysis call and reports a 77% token reduction versus multi-agent approaches.
FAQ
Is AgentSys limited to Claude Code?
Is there a human checkpoint before implementation?
/next-task documents user plan approval before implementation. /prepare-delivery can also be run independently and does not push or create PRs.Which commands can change a remote repository?
/ship can commit, push, create a PR, wait for CI, address review comments, merge, and clean up. GitHub and GitLab workflows require their authenticated CLIs.Does configuration linting need extra setup?
/agnix command requires the separately installed agnix CLI, available through npm, Cargo, or Homebrew.