Agentic Work Systems: designing the system around the Coding Agent

A Coding Agent uses artificial intelligence to read files, change code and use tools. The quality of its work, however, depends on what surrounds it: context, knowledge, rules, environment, state and verification.

This is the scope of the Agentic Work Systems I design: structured local environments for integrating AI into real work, where the agent can understand where it is, what it needs to do, which sources are authoritative, which tools it can use and how to prove that the work has been completed correctly.

A prompt can describe a task. It does not necessarily preserve project decisions, separate stable knowledge from current state, prepare an executable environment or verify what the agent changed.

Integrating a Coding Agent into real work requires several layers:

• readable instructions and boundaries;
• a knowledge base and targeted retrieval;
• tools, CLI, MCP or dashboards;
• operational state and decision logs;
• workflows with inputs, gates and outputs;
• tests, validation and human review.

When these elements work together, AI stops being an isolated request and becomes part of a work system.

I use Agentic Work System to describe the coordinated combination of workspace, knowledge, tools, state, workflows and verification built around a domain.

An individual repository can be an Agentic Workspace. When the system focuses on development, testing and release, it can become an Agentic Development Environment.

My work is about designing these layers and their relationships: not only how to generate an output, but how to make the process repeatable, understandable and controlled.

1. Instructions — role, goal, constraints and source hierarchy.
2. Context — the documents required for the task, without loading everything.
3. Knowledge — stable memory, decisions and verifiable criteria.
4. Tools — scripts, CLI, MCP, APIs, dashboards and executable environments.
5. State — tasks, runs, current status and outputs.
6. Workflow — sequences, gates, escalation and completion criteria.
7. Verification — tests, linting, QA, preflight and human approval.

The implementation changes from project to project. The responsibility does not: give the agent enough autonomy to be useful and enough structure to remain predictable.

I have applied this approach to local systems created for different needs: development, knowledge management, technical marketing operations, project organisation and focused publishing workflows.

The main public example is WP OneShot, a specialised environment for WordPress projects assisted by Coding Agents. It combines isolated local stacks, briefs, instructions, tools, a dashboard and quality gates from preparation to verification.

Alongside WP OneShot, I have built other systems that strengthen my daily work. The domain changes, but the method remains the same: readable context, authoritative sources, suitable tools, explicit workflows and verifiable outcomes.

These systems rely mainly on local files, Markdown, JSON, repository instructions, scripts and open protocols. This file-based foundation makes the architecture usable by different Coding Agents that can access the workspace, read and modify files, use the available tools and follow repository instructions.

In my daily work I use them mainly with Codex and Claude Code. The system is centred on the agent’s operational capabilities, not on the model brand.

The underlying thread has not changed. I have long built digital systems where content, data, interfaces, rules and people need to work together: advanced WordPress, editorial backends, operational workflows, accessibility, tracking and custom tools.

Agentic Work Systems extend the same method to a new kind of operational collaborator. A Coding Agent also needs an understandable structure, reliable sources and controls proportionate to risk.

What is an Agentic Work System?

A coordinated environment of workspace, knowledge, tools, state, workflows and verification, designed so a Coding Agent can contribute to real work with explicit context and controls.

How does it relate to an Agent Harness?

An Agentic Work System can include, configure or extend harness functions such as workspace access, tools, context, state, policies and verification. In my systems this layer integrates with the operational loop provided by the Coding Agent.

Which Coding Agents can use it?

Coding Agents that can access the workspace, read and modify files, use tools and follow repository instructions. The architecture is file-based and does not revolve around a single model; in my work I use it mainly with Codex and Claude Code.

Is it only for software development?

No. The same pattern can support knowledge management, operations, research, publishing and other structured workflows. Agentic Development Environments are the subset focused on development, testing and release.

Does the agent work autonomously?

Only within defined boundaries. The systems include gates, validation and escalation. Responsibility and approval for material actions remain human.

Are you trying to integrate Coding Agents into development or operational work without losing context and control?

We can start from the existing architecture: sources, tools, state, gates and the points where human intervention still matters.

Discuss the system