How Agentic Engineering Automates End‑to‑End Backend Development

Backend developers often lose context when switching between PM, Git platforms, editors, and DevOps tools. The emerging concept of Agentic Engineering addresses this by letting humans define goals, constraints, and quality standards while AI agents execute structured workflows with human checkpoints.

Phase 1: Requirement Capture & Branch Initialization

Tool: pm-dev (Skill) – creates a PM story (or accepts a URL), generates a feature branch feature/{user}_{short_id} , and saves the requirement document under docs/pm/ .

Example command:

❯ /pm-dev RedeemReward
接口里的数据上报逻辑变更，无论领取是否成功，都要上报结果
（上报数据新增 errcode, errmsg 字段）。新字段要更新
git.example.com/org/component/report_data/reportstruct go mod 获取。

The AI detects the current branch, prompts for missing PM URL, and automatically creates the story and branch.

Phase 2: Interactive Requirement Clarification

Tool: superpowers:brainstorming – explores the codebase, asks clarifying questions, and creates a detailed design document saved to docs/plans/ .

AI asks questions such as whether new fields belong to an existing table or a new one, then proposes two implementation options and recommends one.

方案 A（推荐）：在 RedeemReward 末尾统一上报
方案 B：在每个 return 前都加一次上报调用

Phase 3: Planning

Tool: superpowers:writing-plans – reads code details, generates a task list with precise file paths, change descriptions, and verification steps, and saves the plan to docs/plans/ .

Sample plan (4 tasks): upgrade report_data, add fields to CardReportParam, refactor RedeemReward, and run go build / go vet.

Phase 4: Parallel Development Execution

Tool: superpowers:executing-plans (or superpowers:subagent-driven-development ) – runs each task either sequentially or in parallel sub‑agents, automatically generating Conventional Commits.

d81eaee chore: upgrade report_data for new errcode/errmsg fields
18a0975 feat: add ErrCode/ErrMsg to CardReportParam and ReportCard
e978e6b feat: report RedeemReward result regardless of success or failure

Phase 5: Self‑Code Review

Tool: code-review – scans git diff and produces a markdown report with severity levels (Critical, Major, Minor, Suggestion) and categories (security, performance, style, etc.).

│ Critical │ 0 │
│ Major    │ 0 │
│ Minor    │ 1 │   [CR‑可读性] ChannelID 魔法数字缺少注释
│ Suggestion│ 1 │   [CR‑编程规范] Errcode/Errmsg 字段缩进不一致

Phase 6: Compilation & Deployment

Tool: dtools – detects build parameters from Makefile (or generates defaults), performs cross‑compilation, and deploys to the selected environment.

CGO_ENABLED=0 GOOS=linux GOARCH=amd64 go build -v -o .build/reward_service
dtools bpatch -env pre -app my_app -server reward_service -bin .build/reward_service -instances "pre.my_app.reward_service.instance001"

The AI also auto‑updates expired instance names before deployment.

Phase 7: Log Query & Debugging

Tool: galileo-log-query – queries logs with filters (target, namespace, time range, level) and correlates results with code context.

❯ 帮我查一下 pre 环境 reward_service 最近 10 分钟的 error 日志

AI returns the offending line and suggests a fix (e.g., handling nil config).

Phase 8: Merge Request Creation

Tool: /create-mr (command) – extracts the short ID from the branch, fetches the PM title, analyses commits, and auto‑generates a MR title, description, and reviewer list.

标题: RedeemReward 上报逻辑变更，无论成败都上报并新增错误码
## 变更说明
- 升级 report_data v1.0.180 → v1.0.182
- CardReportParam 新增 ErrCode/ErrMsg
- RedeemReward 接口统一上报
## 影响范围
仅影响 RedeemReward 接口的数据上报逻辑

Phase 9: AI‑Assisted Code Review

Tool: /review-mr – loads the MR diff, runs code-review rules, and produces line‑level comments.

repo/report/card_report.go:57‑58 – Minor: Errcode/Errmsg 字段缩进不一致
logic/redeem_reward.go:43 – Suggestion: ChannelID magic number 3 lacks comment

Reviewers decide which comments to submit.

Phase 10: Automated Fixes for Review Comments

Tool: /fix-mr – fetches unresolved comments, generates patches, runs go build and gofmt , then commits with a conventional message.

fix: address CR comments for RedeemReward report changes
- Added comment for ChannelID magic number
- Aligned Errcode/Errmsg indentation

Phase 11: Merge & Release

The MR is merged manually, triggering the CI/CD pipeline that builds the binary, creates a Docker image, and performs a gray‑release. This step remains human‑controlled due to risk considerations.

Overall Token Consumption

The end‑to‑end run consumed a substantial number of tokens (see the token‑usage chart), highlighting the need for higher token quotas in production settings.

Key Takeaways

Agentic Engineering turns AI from a "prompt‑and‑pray" tool into a disciplined, audited participant in the software delivery pipeline.

Human orchestration defines goals and validates each critical hand‑off, while AI handles repetitive, rule‑based tasks such as code generation, commit formatting, MR drafting, and log querying.

Skills and Commands are composable; a single terminal session can drive the entire lifecycle from spoken requirement to production release.

Proper token management and model selection (Claude for code generation, Codex/Gemini for review) improve reliability.