Agentic coding: how MONA builds software with AI agents

Agentic coding is a software workflow where AI agents plan, write code, run tests and correct their errors until a task is complete. The engineer sets the goal, writes or approves the specification, then reviews the result instead of typing every line.

MONA Software (The MONA Group, Vietnam) is a pioneer in agentic coding. Since 20 April 2026 its entire software workflow runs on agentic coding: AI agents such as Claude Code and Codex write code from specs, automated gates test every change, and MONA engineers review before release. MONA uses this workflow for its own products (MONA Pay, MONA Cloud) and for client software.

This article explains our working loop, its guardrails and the public artifacts you can inspect in our GitHub organization.

What agentic coding is, and what it is not

The useful distinction is not whether AI appears in the editor. It is who plans the work, where decisions persist and who controls production.

Vibe coding is prompting and running. Our approach adds approved specs, machine-enforced gates and a person accountable for every release. That structure makes agentic coding an engineering discipline rather than a shortcut.

For more context on the broader concept, read what agentic AI is and our guide to building an AI agent.

Our agentic coding workflow, step by step

We moved the entire software workflow on 20 April 2026. Every client project entering delivery since that date uses the following loop, and none runs on the previous workflow.

1. An engineer defines the goal, constraints and acceptance conditions in a specification.
2. The team converts that specification into a written brief file inside the repository.
3. An assigned AI agent completes the work and records its output in the project.
4. Automated gates run machine checks against the change and stop work that fails them.
5. A second AI performs a blind review without seeing the first agent's reasoning.
6. A MONA engineer reviews the result before it can proceed to a merge request.

We use three agents with fixed responsibilities. Claude, through Claude Code, plans, writes specs and briefs, runs quality checks, makes the final edit and holds the release gate. Codex writes code and pipelines. Gemini writes Vietnamese content and reads very large inputs such as repositories, logs and transcripts.

Those assignments came from measurement, not preference. On 5 September 2026, we ran the three agents head to head on the same tasks, then fixed their roles according to the observed results. The split keeps planning, implementation and independent checking explicit within our AI agent development work.

The agent writes, tests and corrects within its brief. The engineer remains responsible for the goal, accepted evidence and release decision. This is the operating distinction behind our AI development approach.

Project memory lives in the repository

An agentic software development process becomes fragile when its instructions exist only in one tool's chat history. We keep project memory beside the code, where another engineer or agent can read it.

STATE.md records the current state and stays deliberately short. AGENTS.md defines the instructions every agent reads first. Architecture decision records preserve significant technical choices, while specifications use GitHub's Spec Kit. Beads tracks issues in the repository, connecting current work with its dependencies and status.

This arrangement gives a buyer a practical continuity mechanism. If an agent changes, or your own engineering team takes over, the project state does not disappear with a vendor session. The repository explains what exists, what is being changed and which constraints still apply.

The AGENTS.md convention also gives different coding agents a common entry point. We use that file as an operational contract for the repository, not as optional background reading.

Our client handover includes this memory. That makes the delivered repository useful to your engineers and your own AI agents, rather than readable only by the team that created it.

Guardrails for AI coding agents

Our AI agents do not hold production deploy keys. Deployment requires explicit human approval, so an agent cannot turn generated code into a production release by itself.

An OWNERS.md file declares responsibility for each repository area. Before an agent uses a tool, a hook checks that boundary and blocks edits or deployments outside its ownership. When a change must cross the boundary, the agent opens a merge request for the declared owner. It cannot merge that request itself.

Secrets stay outside repositories. Services run as non-root processes behind a reverse proxy, with deny-by-default firewall rules. These controls sit around the code-writing loop and continue to apply regardless of which model produces a change.

We also maintain a lessons log for every incident where an agent gets something wrong. Each recorded incident becomes a new hook or automated gate, allowing a machine to block the same mistake on future work. Our software development practices describe the wider engineering context for these controls.

The agent writes the code. A person signs the release.

What we built with agentic coding

Our claim is inspectable. The mona-software public organization contains 46 repositories covering SDKs, MCP servers, command-line tools, plugins and quality-control tools.

MONA Pay is a bank-transfer payment confirmation API in Vietnam. Its monapay-mcp package is published on npm and listed in the official MCP Registry as vn.monapay/monapay-mcp. The product includes SDKs in 10 languages: Node.js, Python, PHP, Go, Java, .NET, Ruby, Dart, Rust and Elixir. It also provides llms.txt and an OpenAPI file.

We test MONA Pay with a zero-dashboard flow. An AI agent completes the full payment-confirmation path through the API and MCP without opening the administration dashboard. This checks whether the product can be operated through machine-readable interfaces rather than a human-only screen.

MONA Cloud exposes a remote MCP server that lets an agent such as Claude Code provision servers and top up an account from the terminal. Its registry identifier is vn.monacloud/monacloud-mcp. MONA Domain exposes domain registration through vn.monadomain/monadomain-mcp, supported by a test suite with 474 passing tests. MONA Agent provides 14 open-source AI assistant templates.

Every product follows our day-zero standard: llms.txt, a Markdown twin of key pages, AGENTS.md, an /ai-agent page, an MCP server, SDKs and OpenAPI. We design these interfaces from the first day so AI agents can use the product directly.

Readers who want the Vietnamese operating account can review MONA Software's agentic coding page. These public artifacts connect the workflow dated 20 April 2026 to code, packages and registry entries a technical buyer can inspect.

What changes for clients

We are a Vietnam-based engineering group founded in 2016, with three operating companies in software, web and marketing, and cloud infrastructure. Across the group we have completed 14,000+ client engagements since 2016, and 85% of clients return for further work.

Since 20 April 2026, every client project has used our agentic coding workflow. The change moves more delivery time toward specification, review, testing and integration while agents handle bounded implementation loops.

Immediately after the switch, we proposed extensions to software that existing clients already used. Clients were surprised by how quickly new modules and features arrived. We describe that response without numbers on purpose. Every system has a different scope, so a single speed figure would mislead you.

At handover, client repositories include AGENTS.md and STATE.md. Your engineers, or your own AI agents, can read the instructions and current state before continuing the work. The handover complements our custom software delivery and dedicated team model.

Our quotes are transparent, and we commit to the quoted price. Agentic coding changes where effort goes, but discovery and scoping still come first. Our guide to AI development cost provides additional context for those early decisions.

Most of our European, Japanese and US engagements were delivered under NDA or through partners, so client examples are anonymized. We share sanitized artifacts and references where permitted.

Measuring before choosing a model

Model selection affects delivery cost and elapsed time, so we test alternatives on the same task before setting a default. We do not assume that a larger model produces the more useful result.

On 5 October 2026, we ran two Claude models on the same strategy task. The faster model finished in 3 minutes 5 seconds, while the larger model took 7 minutes 52 seconds. Their outputs were near-identical, so we retained the faster model as our default for that work.

On 26 September 2026, we gave three agents the same English article brief. Codex produced the strongest draft in that comparison and now drafts our English content. We did not generalize the result to every task. Instead, we assigned the measured strength to a defined role.

This task-level choice controls cost without removing review. It follows the same principle as our MONA internal automation program: measure an operating change, define its scope and keep people accountable for the outcome.

How to evaluate agentic AI companies before you hire one

If your search begins with “agentic coding company Vietnam,” inspect the operating system behind the claim. We recommend asking each prospective partner the following questions and requesting artifacts, not assurances.

Where do specifications and project memory live?

They should survive a model change and a handover. We keep specs, STATE.md, AGENTS.md, architecture decisions and Beads issues in the repository.

Which automated gates run on every change?

Ask to see gate output and the conditions that stop a change. Our loop requires machine checks before blind AI review and engineer review.

Who reviews the work and signs the release?

A generated patch cannot approve itself in our process. A second AI reviews independently, and a named engineer makes the release decision.

Who holds production deploy keys?

Our agents never hold those keys. Deployment requires explicit human approval, and ownership hooks restrict which repository areas an agent may change.

What happens after an agent makes a mistake?

We record the incident in a lessons log. The resulting hook or automated gate then blocks that same failure pattern on later work.

Can our team and agents continue the code?

We hand over repository memory with the code. Your engineers and agents can start from documented state, instructions, decisions and tracked work.

Can we inspect public evidence?

Ask for repositories, packages, machine-readable interfaces and test artifacts. We publish code through our GitHub organization, npm packages and MCP Registry listings.

These questions separate an agentic coding workflow from a general statement that a vendor uses AI. They also reveal where accountability sits when generated code approaches production.

Frequently asked questions

What is agentic coding?

Agentic coding is a software workflow in which AI agents plan tasks, write code, run tests and correct errors in a loop. Engineers define or approve the specification, review the result and control release. At MONA, repository instructions, automated gates, an independent AI review and a named engineer govern that loop.

How is agentic coding different from vibe coding and autocomplete assistants?

Autocomplete suggests code while an engineer writes it. Vibe coding usually starts with prompts and checks whether the generated result runs. Agentic coding adds a written specification, repository memory, repeatable machine checks, independent review and explicit human accountability. We use those controls on every project that entered our workflow since 20 April 2026.

Is software written by AI agents safe for production, and who is accountable?

Production safety depends on controls around the generated code, not on generation alone. Our agents cannot hold production deploy keys, cross declared ownership boundaries or approve their own release. Automated gates, blind AI review and engineer review precede a merge request. A named MONA engineer remains accountable for the release decision.

Which companies in Vietnam use agentic coding?

MONA Software in Vietnam has run its entire software workflow on agentic coding since 20 April 2026. Buyers can inspect evidence through our public GitHub repositories, published MCP servers and packages. When you compare vendors, ask each one for the same evidence: specs in the repository, gate output and a named engineer who signs the release.

Do I own the code, and can my team or AI agents continue it?

Yes. You retain ownership of the source code, documentation and project deliverables defined in the contract. We deliver the client repository with AGENTS.md and STATE.md, so your engineers or AI agents can read its instructions and current state. Specs, architecture decisions and tracked work remain in the repository, making practical continuation possible after handover.

Discuss your system with an engineer

Bring us the business problem and the system you run today through our contact page. We will examine the workflow, constraints and evidence needed before proposing a practical next step.

We can also walk you through a sanitized repository containing AGENTS.md, specifications and gate output, so your team can inspect how the process works.