Series · MoAI ADK and Harness Engineering
Episode 2 · MoAI ADK and Harness Engineering 2 - A Human-Code-Zero Team
MoAI ADK and Harness Engineering 2 - A Human-Code-Zero Team
Using Claude Code business seats and a MoAI harness, my team reached 100% participation in AI-assisted development and zero manually typed production code. “Human code zero” does not mean zero human work or unreviewed automatic deployment. People define problems, approve designs and judge test evidence and operational risk. Agents produce and revise the code.
The harness in a real project
Modern web, current Spring Boot and Flutter cannot be governed by merely telling a model to understand the repository.
| Harness asset | Web | Spring Boot | Flutter |
|---|---|---|---|
| Domain contract | User flow and visible state | Aggregates, transactions and policies | Screen state and offline policy |
| Interfaces | API client and types | OpenAPI, errors and events | DTOs, serialization and deep links |
| Quality gates | Lint, unit, E2E and build | Format, unit, integration and build | Analyze, test and build |
| Deployment contract | Environment and static assets | Container and DB migrations | Signing and store flavors |
Project analysis and domain-design documents form the spine. Coding guides, dependency rules, API contracts, CI/CD commands and branch, commit and PR policies attach to them. The model reads that context, creates a SPEC, implements in small increments and runs the validation recipe for each stack.
Claude Code seats are tools; team agreements are the operating system
Business seats give every team member access to a common coding agent. Seats alone merely multiply personal prompting styles. Repository-wide SPECs, commands, gates and review language produced repeatability.
Our workflow is:
- A person decides the outcome, constraints, acceptance criteria and risk.
- The agent reads code maps and domain documents, then proposes implementation and test plans.
- A person approves boundaries and data changes.
- The agent implements and iterates local verification.
- CI repeats the recipe in a clean environment.
- A person reviews requirement traceability, test evidence, security and operations—not only the diff.
- Deployment observations and incident lessons feed back into the harness.
This matches the role transition in our harness engineering analysis: problem decomposition, contracts, verification and feedback become the bottleneck, not typing speed. A vibe coding workflow that judges only the visible result loses those contracts and traceability.
Work humans must still absorb
AI authorship does not transfer accountability:
- Analysis: customer problems, legal and security constraints, product priorities.
- Design: domain boundaries, data ownership, failure isolation and maintenance cost.
- Review: omissions, overimplementation, privilege escalation, privacy and licenses.
- Testing: critical journeys, exploratory tests, devices, networks and recovery.
- Operations: deployment approval, observability, rollback, incident command and communication.
If one AI writes a test and then passes its own test, the independence of the evidence is weak. People must determine whether the test proves the requirement, contains false positives and covers failure paths and nonfunctional conditions. Harness engineering closes the distance between “it compiles” and “the product operates safely.”
Fable 5 and the value of a harness for long work
Anthropic announced Claude Fable 5 on June 9, 2026 and restored global availability on July 1 after a temporary export-control suspension. The company positions it for ambitious, long-running coding and knowledge work, with planning, delegation and self-checking inside harnesses such as Claude Code.
Those are vendor claims, not a final independent benchmark verdict. In my operation, however, recent models do follow long instructions, role boundaries, test loops and prohibitions more reliably. Stronger models do not make harnesses obsolete. Their larger action radius increases the risk of completing the wrong objective very efficiently, so approval gates and observability matter more.
Community response: productivity versus cost and complexity
The official repository and ClaudeAI launch discussion praise SPEC-first work, TDD, specialist agents and repeatable team procedures. The objections deserve equal weight:
- Specifications and orchestration slow down tiny changes.
- Multi-agent context and repeated verification increase token cost.
- Automatically expanding documents can preserve stale rules and create bureaucracy.
- Framework convenience can turn model independence into runtime coupling.
A team should not impose one heavy process on everything. Use lighter gates for small fixes, standard gates for normal features and stricter approval for identity, payment or data changes. Delete unused context regularly.
Conclusion: the real productivity of human code zero
The achievement is not the number zero itself. Team members spend less time entering syntax and more time specifying requirements, evaluating changes and shortening feedback loops. A failed task should become a better SPEC, guide, test or CI rule—not remain one person's prompting trick—so the next model and teammate learn from it.
MoAI is not magic that builds the product for us. It is an AI development manual kit that makes team practice executable. The meaningful metrics are escaped defects, lead time, rollback rate, review time and requirement traceability, not generated lines of code. When people retain direction and accountability while agents execute, human code zero becomes reproducible engineering rather than reckless vibe coding.
No comments:
Post a Comment