agentcorp

AgentCorp

Loop engineering for software delivery: controllable, understandable, verifiable.

English · Simplified Chinese

Quick Start · Skills · Artifacts

AI can generate code faster and faster, but the cost of verifying whether that code is correct still lands on you. When you receive a pile of code that "looks fine," the responsibility for judging whether it is actually right is entirely yours. The deeper your dependency on AI becomes, the easier it is for that judgment to get dull.

The harder problem is the loop that follows: an agent's work process is a black box, so you cannot understand its reasoning; because you cannot understand it, you skip review; because you skip review, cognitive debt accumulates; the deeper that debt gets, the more dependent you become, and the harder it is to correct the agent when it goes wrong. Eventually, you no longer dare to hand it important work.

AgentCorp exists to break that loop. It is a loop engineering system for software delivery: it turns agent work from an uncontrollable, unreadable, untraceable black-box chain into a controllable, understandable, verifiable delivery loop. It includes 32 skills drawn from enterprise-grade software delivery practice, covering the full delivery loop and the supporting practices needed to run it in real projects. It works with both Claude Code and Codex.

• Controllable -- The process scales itself to the size of the task. A one-line change does not pay the cost of an architecture review, while a new system skips no critical phase. Gates actually block: failed verification stops the pipeline, and repeated failure forces replanning. You can step in at any point, or let the pipeline run.
• Understandable -- Every phase leaves a structured artifact and records who made which decision under what evidence. Every review finding is explained to the level of "even if you have not read this code, you can still judge whether it should be changed." At delivery time, the final diff is turned into function-level comments that explain why the change was made.
• Verifiable -- No role can approve its own output. Tests are decided before implementation and independently reviewed. Review findings are treated as possible false positives and re-verified by another role; only confirmed findings enter the fix stage.

Quick Start

Install

Claude Code:

/plugin marketplace add ylxmf2005/AgentCorp
/plugin install agentcorp@agentcorp

Then run /reload-plugins or restart. Skills are namespaced, for example /agentcorp:delivery-orchestrator.

Codex:

codex plugin marketplace add ylxmf2005/AgentCorp

Launch Codex, enable AgentCorp from the /plugins menu, and restart. To install a single skill: use skill-installer to install the skill at repo ylxmf2005/AgentCorp path agentcorp/delivery-orchestrator.

Calling Skills

In Claude Code, call a skill as a slash command:

/agentcorp:explain output_mode=artifact explain review/code-review.md for a sponsor

In Codex, ask for the skill by name or $skill-name:

Use $explain with output_mode=artifact to explain review/code-review.md for a sponsor.

If you omit options such as output_mode, the skill uses its default behavior.

First Use

After installation, describe what you need, or call /agentcorp:delivery-orchestrator. It will first confirm the success criteria with you, recommend an execution route, then move through the phases and stop at each gate to report.

When a task needs a real browser or logged-in state verification, AgentCorp uses an isolated browser profile, asks you to log in manually once, and then performs checks inside the page automatically. It does not touch your local cookies. At the end of the task, you receive a delivery report and an audit record that traces every decision.

Skills

The 32 skills are grouped by function below. Each skill's behavior is defined in agentcorp/<skill>/SKILL.md and appears in the Claude Code and Codex skill pickers. Together, they cover the delivery loop and the supporting behaviors needed to run it in real projects.