Agentic Software Development Life Cycle

The HyperShift team operates an Agentic Software Development Life Cycle (ASDLC). A framework that decouples distributed consumption of reusable agentic building blocks (skills, slash commands, sub-agents, knowledge bases) from the systems that orchestrate agent execution and LLM inference at scale.

This separation enables flexible implementation choices: the same building blocks power local Claude Code (or other agents) sessions that can run on a developer's laptop, centralized Prow CI jobs, ambient agents, long-running cluster workloads, etc.

graph TB
    subgraph "Building Blocks Layer"
        direction LR
        SK["Skills & Slash Commands — /jira:solve, /code-review:pr, ..."]
        AG["Sub-agents — HCP SMEs, Explore, Plan"]
        KB["Knowledge Bases — CLAUDE.md, AGENTS.md, Rules"]
        HK["Hooks & Plugins — Pre-commit, CI integration"]
    end

    subgraph "Orchestration / Execution Layer"
        subgraph Local["Local (Dev Laptop)"]
            direction LR
            L_CC["Claude Code CLI"]
            L_BM["Bot Minter"]
        end
        subgraph Centralized["Centralized (Shared Infra)"]
            direction LR
            C_PJ["Prow Jobs"]
            C_AM["Ambient"]
        end
    end

    SK --> Local
    AG --> Local
    KB --> Local
    HK --> Local
    SK --> Centralized
    AG --> Centralized
    KB --> Centralized
    HK --> Centralized

    style SK fill:#e8f4fd,stroke:#1a73e8
    style AG fill:#e8f4fd,stroke:#1a73e8
    style KB fill:#e8f4fd,stroke:#1a73e8
    style HK fill:#e8f4fd,stroke:#1a73e8
    style L_CC fill:#fce8e6,stroke:#d93025
    style L_BM fill:#fce8e6,stroke:#d93025
    style C_PJ fill:#fce8e6,stroke:#d93025
    style C_AM fill:#fce8e6,stroke:#d93025

Phases of ASDLC

Phase 1 — Discovery

The discovery phase focuses on understanding the problem space, designing the solution, and refining scope into actionable work items.

flowchart LR
    subgraph "Architecture Design"
        A1["Interactive Claude session: discuss design and implementation"] --> A2["HCP SME sub-agents: cloud-provider, control-plane, data-plane, api, architect"]
        A2 --> A3["Create RFC or enhancement proposal"]
    end

    subgraph "Scope Refinement"
        B1["/jira:create — Create Jira issues"] --> B5["/jira:ready-to-solve — Validate readiness for agent"]
        B1 --> B2["/jira:generate-test-plan — Generate test plan from Jira"]
        B1 --> B3["/utils:generate-test-plan — Generate test plan from PRs"]
        B1 --> B4["Obra superpowers"]
    end

    A3 --> B1

Architecture Design

Scope Refinement

Phase 2 — Delivery

The delivery phase covers implementation, review, and quality assurance — each supported by agentic workflows.

flowchart LR
    subgraph "Implementation"
        C1["/jira:solve"]
        C2["Bot minter"]
    end

    subgraph "Code Review"
        D1["CodeRabbit — automated review"] --> D3["Address feedback"]
        D2["/code-review:pr — /code-review:pre-commit-review"] --> D3
    end

    subgraph "QE & Testing"
        E2["/ci:analyze-prow-job-test-failure"]
        E3["/agentic-qe test-plan"]
    end

    C1 --> D1
    C2 --> D1
    D3 --> E2

Implementation

Code Review

QE & Testing

Modes of Agentic Workflows

A mode describes where/how execution happens. Building blocks can be executed in different modes depending on the use case:

graph LR
    subgraph "Local"
        L1["Developer laptop — Claude Code CLI"]
    end

    subgraph "Centralized"
        P1["Prow periodic jobs"]
        P2["Prow presubmit jobs"]
        AM1["Ambient agents"]
        LR1["Long-running agents — prebuilt binaries as cluster workloads"]
    end

    L1 -.- BB["Shared — Building Blocks"]
    P1 -.- BB
    P2 -.- BB
    AM1 -.- BB
    LR1 -.- BB

    style BB fill:#e8f5e9,stroke:#34a853

Local

Building blocks executed by developers on their laptops using Claude Code CLI, Bot-Minter, others... This is the most interactive mode where removing ambiguity might need several iterations. Ideal for design discussions, exploratory work, and ad-hoc tasks.

Before publishing a code artifact for human review, developers are expected to levereage building blocks for local code review and agentic qe

Centralized

Building blocks run and driven by central infrastructure tools. These run on schedule, in response to events, or continuously as cluster workloads: Prow Jobs, Ambient, Cluster workloads... Ideal for repetitive tasks or concise work items.

Hands-off Delivery

flowchart TD
    A["Jira item labeled with issue-for-agent"] --> A2["/jira:ready-to-solve validates readiness"]
    A2 --> B["Periodic Job picks from backlog"]
    B --> C["Agent generates code, reviews, addresses and creates PR"]
    C --> D["Agent review on published PR"]
    D --> E["Human review"]
    E --> G["Human lgtm"]
    G --> H["/agentic-qe"]
    H --> I["Presubmit e2e runs"]
    I --> J["Auto-analysis on failures"]
    J -->|"Failures found"| E
    I -->|"Pass"| K["Merge"]

For details on the current centralized jobs, see AI-Assisted CI Jobs.

Monitoring

Dashboards

Agent execution metrics, token usage, and cost tracking

Slack Integration

Building blocks integrate with Slack for notifications, status updates, and interactive assistance.

Principles

• Reuse drives self-improvement — Every execution of a building block generates signal (successes, failures, review feedback) that feeds back into refining the building blocks themselves.
• Raise the quality floor — Agentic workflows must enforce consistent code patterns, code review, test generation, and CI analysis on every change, raising the baseline quality across all contributions.
• Compounding returns — Better building block specifications produce better output, which produces better input data for the next iteration. The system improves itself over time.

Getting Started

/plugin marketplace add openshift-eng/ai-helpers
/plugin install jira@ai-helpers
/plugin install utils@ai-helpers
/plugin install ci@ai-helpers
/plugin install code-review@ai-helpers