Introduction¶

The most common problem in multi-agent teams is not insufficient model capabilities, but inconsistent rules: DRI, reviewer, lead, and ops write their own instructions separately, copy the same handoff format four times, and change one spot but forget the other three. agent-loop-workflow is a project-agnostic multi-agent collaboration workflow skeleton. It converges role responsibilities, handoff formats, review routing, and delivery order into a single skill. Each agent simply loads the same skeleton.

Participants are plugins within the DSH (DeepSeek Harness) ecosystem. The maintainer is LeslieWylie. License is MIT.

What is it¶

One-sentence positioning: A project-agnostic multi-agent collaboration workflow skeleton for agent teams composed of DRI + reviewer + lead + ops.

It solves the problem of avoiding the need to copy the same rules into every agent’s instructions. The skeleton only defines the collaboration process itself, decoupled from specific projects, issue trackers, and code hosting platforms.

Core Features¶

Below is an introduction to the eight modules listed in the README, broken down in the order of “Role — Constraint — Handoff — Review — Closure”.

Role Topology: Defines four types of roles—DRI (executor), reviewer (evaluation), lead (decision/routing), and ops (automation)—along with their responsibility boundaries.

Loop Guard Invariants: Turn limit, no-progress detection, same-error circuit breaker, time cost warning, explicit exit gate, single authority write. This is the layer that prevents agents from spinning idle or retrying repeatedly.

Standard Handoff Format: Six-field handoff template—Owner / Goal / Input / Write Operation / Acceptance Criteria / Failure Evidence.

Risk Tier Routing: Fast / Standard / High tiers, automatically routing to different reviewers based on the scope of changes.

Delivery Sequence: Verification → commit → push → Draft MR/PR → Readiness Signal → in_review. A fixed process.

Review → Closure Protocol: Automatic dispatch after in_review, reviewer conclusions given as metadata, supporting retry and escalation mechanisms.

Anti-Loop: Idempotent issue creation, state changes not triggering duplicate dispatch, manual source identification.

General Red Lines: Key confidentiality, token permission separation, confirmation for destructive operations.

Installation and Activation¶

The installation command is taken from the original README. Note that the repository ownership is LeslieWylie/agent-loop-workflow:

dsh plugin --profile web add "github:LeslieWylie/agent-loop-workflow"

After restarting, load the skill in any session:

load agent-loop-workflow

The two steps serve respectively: plugin installation registers the skill file into the current DSH profile; the load command enables it within the session, after which agents can reference the rules in the skeleton.

Typical Usage¶

After the above steps, a typical turn can proceed as follows:

1. After the DRI completes a task, they hand off their output to the next step using the six-field handoff format.
2. The Lead determines the risk tier (fast / standard / high) based on the scope of changes and routes it to the corresponding reviewer.
3. The Reviewer provides a conclusion as metadata during the in_review phase; if not passed, it is sent back for retry or handled via escalation.
4. Ops handles automation actions; the Loop Guard invariants continuously monitor turns, progress, and time cost.

The entire process does not depend on external services, only invoking DSH’s skill loading mechanism. The README also explains that the plugin is not bound to a platform: simply map “issue” and “MR” to your own system.

Use Cases and Considerations¶

Suitable for scenarios where you want to unify collaboration standards for an agent team but don’t want to repeatedly copy rules into every agent’s instructions. The three design principles given in the README are:

• Project-Agnostic: Only defines the workflow skeleton. Specific project code pitfalls, checklists, and repository paths should be placed in each project’s own *-conventions / *-engineering / *-review-rules skill.
• Zero Dependencies: Requires no external services, only relies on DSH’s skill loading mechanism.
• Platform-Agnostic: Does not specify an issue tracker or code hosting platform.

Two points need attention. First, the plugin runs with the permissions of the current DSH process; before installation, check the source code and license (this project is MIT). Second, package.json shows version 0.1.1, Node engine requires >=18, and devDependencies includes @deepseek-ai/cordis and @deepseek-ai/dsh-skill, indicating it uses the DSH cordis patch bundle mechanism. Before making changes, it is recommended to run the built-in test scripts first:

npm test
node scripts/check-pack.mjs

Conclusion¶

The value of agent-loop-workflow lies in converging repetitive and error-prone process rules from multi-agent collaboration into a shareable skeleton, rather than copying instructions everywhere. If your team is building a DRI + reviewer + lead + ops team, you can start with this role topology and six-field handoff format.

Directory Page: https://www.skillhub.cn/plugins/YaoleWu/agent-loop-workflow
GitHub: https://github.com/LeslieWylie/agent-loop-workflow

(Note: YaoleWu in the directory page path does not match the actual ownership of the README and GitHub repository LeslieWylie; the repository content takes precedence.)