AI Development as a System for Delivering Changes Safely

This guide organizes what Netsujo has learned from operating AI-assisted development in practice. It connects Git and GitHub, pull requests, CI, Agent Harness design, state management, authority, independent QC, recovery, deployment, and production verification into one delivery model.

The full delivery flow

Stage	Working unit	What must be made explicit
Request	Issue	Purpose, constraints, and acceptance conditions
Change	Commit / Pull Request	A reviewable change with history
Verify	CI / Review	Which exact version was checked and what passed
Publish	Deploy	Which verified artifact or commit reached the target environment
Observe	Monitoring	What is actually running and visible in production

Six operating principles

Track implementation, verification, publication, and production observation as separate completion states.
Bind verification evidence to the exact version that was checked.
Treat the model, the agent, and the surrounding execution harness as different layers.
Track current state, execution authority, and verification evidence separately.
Use independent QC so implementation assumptions are checked from a separate runtime and context.
When an external outcome is uncertain, read the current state before retrying.

The operating stack

Git / GitHub / Actions: The history, collaboration, and automated-verification layer: commits, branches, pull requests, review, CI, and deployment evidence.
Agent Harness: The execution environment around an agent: instructions, tools, context, execution loops, guardrails, observability, and recovery.
Agent OS: Netsujo’s operating rules for state, authority, and evidence. This is a Netsujo operating definition, not an industry standard.
Controller: The Netsujo role that tracks current state, writer ownership, progression conditions, and whether execution is allowed to continue.
Orchestrator: The Netsujo role that coordinates assignment, parallel execution, recovery, and independent verification across multiple runtimes.

Start with three core guides

Agent HarnessUnderstand the execution environment that connects models, tools, context, guardrails, observability, and recovery.

State, Authority, and EvidenceSeparate current state, execution rights, and verification evidence across pull requests, deployment, and production.

Independent QCSeparate implementation from verification and accept an exact target from a different runtime and context.

Choose a three-level learning path

LEVEL 1: Read the current state in GitHub first

Separate branch, HEAD, base and head, pull request, checks, merge, and deployment so you can tell what changed and what was actually verified.

Parallel AI-agent development with Git worktrees Fix the verification target with an exact SHA

LEVEL 2: Design AI as an execution system

Define agent roles, the harness, stop conditions, state retention, and authority boundaries around the model.

The Agent Harness model Role separation across ChatGPT, Claude Code, and Codex

LEVEL 3: Add governance and recovery for production delivery

Use state, authority, evidence, independent QC, recovery, and the Netsujo Agent OS operating model to reach a safe terminal state.

Separate State, Authority, and Evidence Independent QC acceptance conditions Netsujo Agent OS in operation（日本語）

Read the knowledge map by problem, not chronology

Foundation / Execution: Role separation, MCP and tool design, RAG, specifications, Git worktrees, and parallel execution.
State / Governance: State, authority, evidence, CI, review, approval, deployment, production verification, and recovery.
Incident / Failure: Real incidents converted into rules, policy, gates, regression tests, and operating contracts.
Evidence / Evolution: First-hand records of what changed in practice, what became faster, and where human judgment remained.
Adoption / Business: AI implementation diagnosis, business-tool connection, RAG, and the path from operating knowledge to implementation support.

Knowledge Map: jump to a lane

Foundation / ExecutionRole separation, specifications, tools, MCP, RAG, Git worktrees, and the execution environment around AI agents.

State / GovernanceState, authority, evidence, CI, review, approval, deployment, production verification, and recovery.

Incident / FailureReal incidents converted into rules, policy, gates, regression tests, and operating contracts.

Evidence / EvolutionFirst-hand records of what changed in practice, what became faster, and where human judgment remained.

Adoption / BusinessDiagnosis, applicability, roadmap, implementation support, and the business path into AI adoption.

Foundation / Execution

Role separation, specifications, tools, MCP, RAG, Git worktrees, and the execution environment around AI agents.

Design with ChatGPT, implement with Claude Code, audit with CodexPrimary core guide: Independent QC.Run AI agents in parallel without breaking the repositoryPrimary core guide: State, Authority, and Evidence.Connect business tools through MCPPrimary core guide: Agent Harness.Run RAG with business knowledgePrimary core guide: Agent Harness.Specification-driven development with Claude CodePrimary core guide: Agent Harness.

State / Governance

State, authority, evidence, CI, review, approval, deployment, production verification, and recovery.

Incident / Failure

Real incidents converted into rules, policy, gates, regression tests, and operating contracts.

AI Agent Development and Operations Incident LogSeries hub; no forced primary core guide.The chat name was not the work-item identityPrimary core guide: State, Authority, and Evidence.A misimplementation in AI-agent SEO improvementPrimary core guide: Independent QC.

Evidence / Evolution

First-hand records of what changed in practice, what became faster, and where human judgment remained.

The company got faster with AI. Did the founder get time back?（日本語）Evidence record; no forced primary core guide.One month of running four projects in parallel with Claude CodePrimary core guide: Agent Harness.Solo development and the loneliness of judgment in the AI-agent era（日本語）Evidence record; no forced primary core guide.

Adoption / Business

Diagnosis, applicability, roadmap, implementation support, and the business path into AI adoption.

How an AI implementation diagnosis worksAdoption guide; no forced technical core-guide assignment.Where to discuss AI development and adoption in KyotoRegional adoption guide; no forced technical core-guide assignment.

Browse the full Tech BlogThe Knowledge Map above contains the 25 AI-development articles currently organized by Netsujo. The broader Tech Blog also covers Web3, SEO, AIO, implementation, and other engineering topics.

Frequently asked questions

What becomes difficult first in AI-agent development?: Before model capability becomes the limiting factor, teams often struggle to track who is changing what, which version was verified, and how far a change has reached. State, authority, evidence, and recovery become more important as the number of agents grows.
What is an Agent Harness?: In this guide, Agent Harness is Netsujo’s umbrella term for the surrounding execution environment: instructions, tools, context, execution loops, guardrails, observability, and recovery. Other products and organizations may draw the boundary differently.
If CI is green, is the change ready for production?: No. CI proves that defined automated checks passed for a particular target. Merge, deployment, and production observation are separate stages with separate evidence.
Are the Controller and Orchestrator the same thing?: Netsujo separates them. The Controller tracks current state, progression conditions, and execution authority. The Orchestrator coordinates assignment, parallel execution, recovery, and independent verification. This is a Netsujo operating model.
How is Independent QC different from ordinary code review?: Netsujo emphasizes a separate runtime and context checking the same exact target, together with tests, evidence, and operational risks. When the target changes, the previous PASS is not silently reused.
How do this guide, Agent OS, and the AI Practice Community differ?: This guide is the permanent learning hub. Agent OS is the operating proof and control design used by Netsujo. The AI Practice Community is a place for continued exchange around events and practical lessons.

See Netsujo Agent OS in operation（日本語）The practical control-plane work that applies these operating principles inside Netsujo.

Netsujo applies these lessons to PoC acceptance criteria, multi-agent role design, quality gates, authority boundaries, recovery, and production observation.

Discuss AI implementation and development operations