From chatbot to agent harness: how my AI workspace evolved

I started with GPT-3.5 in a browser window. Today I use AI systems that search repositories, edit files, run tests, structure meeting notes, prepare proposals, and turn loosely worded ideas into concrete tasks.

Only a few years separate these two ways of working. For me, however, the decisive progress was not simply moving from one language model to the next. The real leap came when the model became part of an agent harness.

By that I mean a working environment that gives the model context, provides tools, lets it perform actions, and makes the results verifiable. This is what turned AI from an interesting chat window into a daily production tool for me.

I have now worked with several of these environments: GitHub Copilot, Cursor, Claude Code, and Codex. I no longer use some of them, use others daily, and reserve others for particularly difficult tasks. Above all, this taught me what we must contribute ourselves to make these tools productive.

A language model initially generates text. It can explain, summarize, develop ideas, and suggest code. That is useful, but it does not automatically know my project, my way of working, or the intended outcome. Nor can it independently verify whether a suggestion actually works.

An agent harness builds a working environment around the model. In my practical definition, it has five parts: context from files, documentation, rules, and decisions; tools such as repositories, terminals, browsers, and interfaces; execution of changes within clear boundaries; feedback from tests, errors, diffs, and reviews; and guardrails for permissions and approvals.

Anthropic now uses the term itself and describes Claude Code as an agent harness for long running, multi-step work. I find the term useful because it shifts attention from the model to the complete work system. Source: Anthropic on Claude Fable 5

My starting point was GPT-3.5 on the OpenAI website. As with many people, it began as a chat: enter a question, read the answer, ask a follow-up, and copy the text.

It was impressive, but responsibility for the entire workflow remained with me. I had to select the necessary context and paste it into the chat, transfer answers into my tools, and verify whether they were factually correct and practically useful.

Even so, this phase mattered. I learned to phrase tasks in ways a language model could work with. I also learned quickly that a convincing answer is not yet a reliable result. The chat was a good conversation partner, but it was not yet an integrated contributor to a workflow.

The next step was GitHub Copilot. For the first time, I did not have to move all the context from my development environment into a separate chat window. AI appeared where the code was already being written.

Copilot mainly helped me develop faster. Suggestions appeared directly in the editor, repetitive code was quicker to write, and individual questions could be clarified in the immediate project context. GitHub describes Copilot as a coding assistant that can provide code suggestions, answer questions, and prepare changes. Source: GitHub Docs

For me, integration was the decisive factor at the time. The model itself was no longer the only source of value. Its position in the workflow became at least as important. Yet the basic division of labor remained: Copilot suggested, while I implemented and checked every step.

About a year and a half ago, someone recommended Cursor to me. I was skeptical at first. I already had an editor and an AI assistant, so why add another tool?

The answer soon became clear. Cursor did not treat AI merely as an addition for individual lines of code. The agent could understand broader relationships in the repository, modify several files, and implement complete features from a task description. Source: Cursor Docs

This also changed my role. I wrote fewer individual lines myself, but I had to formulate requirements more clearly, keep the architecture in view, and review generated changes more systematically. Agentic software development did not mean thinking less. It shifted more of the thinking from manual input to specification, steering, and review.

Claude Code was the first of these tools that I used consistently from the command line. This move initially seemed unremarkable, but it was the next major leap in my workflow.

In the terminal, the agent could do more than read and change code. It could run commands, start tests, evaluate errors, and continue working. Anthropic describes the CLI in exactly these terms: editing files, running commands, and managing a project from the command line. Source: Claude Code Docs

This created a feedback loop: understand the task, inspect the project, plan and implement the change, run tests, analyze errors, and prepare the result for my review. This process is much more valuable than a single good code suggestion because the agent can observe the consequences of its own work.

At the same time, permissions and guardrails became especially important. An agent with terminal access can work more productively, but it can also turn incorrect assumptions into real changes more quickly. Controlled access, small work packages, tests, and a final review therefore belong together for me.

Today I use Codex both from the command line and in the desktop app. For my personal usage pattern, Codex offers a compelling balance of price and performance. Price alone, however, is not the deciding factor.

What matters is that I can adapt the environment to my actual work. OpenAI documents Codex capabilities including rules through `AGENTS.md`, skills, plugins, MCP connections, terminal work, and local or cloud environments. Source: OpenAI Developers

In my setup, Codex is connected to an LLM wiki built around the Karpathy pattern. It is not a loose collection of notes, but a maintained knowledge context: working principles, decisions, project status, roles, and recurring processes are structured so the agent can take them into account for a specific task.

That is relevant beyond developers. Executives and managers can use the same principle to build a personal or organization-wide working memory. Decisions remain traceable, meetings lead to concrete next steps more quickly, and recurring management tasks do not start from zero every time. This requires maintained content, clear access rights, and professional review of the results.

Skills then encode recurring workflows. Connectors bring the agent to where the actual work happens. The wiki provides persistent context, while skills and connectors turn it into concrete, verifiable work steps.

As a result, I no longer use Codex only for programming. It also helps me structure meeting notes, prepare proposals, turn rough ideas into concrete work packages, summarize project status, and standardize recurring processes through skills.

The important point is not that Codex completes every one of these tasks without review. Its value comes from giving the agent the necessary context and suitable tools in a shared working environment. I do not have to start from zero for every task.

For me, a productive AI stack does not need a single winner. Codex is my daily companion. For particularly difficult tasks and complex relationships, I still use Claude Code selectively, now also with Claude Fable 5.

Anthropic positions Fable 5 as its most capable generally available model for demanding, long running coding and knowledge work. That is a vendor claim, not an independent comparison. It does, however, fit how I divide the roles: I use the model where the additional depth justifies the higher effort for me. Source: Anthropic

This specialization is more useful to me than trying to solve every problem with the same tool. Some tasks need speed and close integration into the daily workflow. Others need especially broad context, long reasoning paths, or a second independent review.

1. Context beats the perfect individual prompt. When an agent knows project rules, architecture decisions, and existing documentation, I have to explain and correct less. An LLM wiki, clear repository rules, and current specifications are therefore part of the productivity infrastructure.

2. Tool access changes the quality of work. A model that only answers can provide a plausible suggestion. An agent that can read files, run commands, and check the result receives feedback from reality. This does not make it automatically correct, but it makes errors more observable and correction loops shorter.

3. Good specification becomes more important, not less. The more autonomously an agent works, the more costly an unclear task becomes. I therefore invest more time in goals, boundaries, acceptance criteria, and verification. With a clear specification, the choice between suitable harnesses is often secondary.

4. Execution and validation belong together. In coding, that means tests, static checks, diffs, and reviews. For proposals or meeting notes, it means sources, known facts, templates, and my professional approval. The exact validation method differs, but the principle remains the same.

5. Productive use is an operational discipline. Which data may the agent see? Which tools may it operate? When must it ask? Which actions require approval? How is the result checked? Which knowledge remains available? This is where an impressive demo becomes a dependable workflow.

GPT-3.5 on the OpenAI website introduced me to working with language models. GitHub Copilot brought AI directly into the editor and accelerated code creation. Cursor was the formative intermediate step toward agentic implementation of larger features.

Today I use Claude Code selectively for demanding tasks, long work chains, terminal execution, and tests. Codex is my adaptable harness with a wiki, skills, and integrations, making it my daily companion for development, knowledge, and project work. Claude Fable 5 in Claude Code gives me additional capacity for especially complex relationships.

This division is not a universal ranking. It describes my current workflow. Different tasks, security requirements, or enterprise environments may lead to a different selection.

My current AI stack is powerful, but it still has one practical limitation: a substantial part of the harness runs on my local computer. Its availability therefore depends on whether that device is powered on, reachable, and configured correctly.

My next step is to move the harness to a VPS. The agent should run continuously and remain accessible from anywhere. Tasks could continue without my laptop serving as the central execution environment, while knowledge, skills, and recurring workflows would come together in one controlled place.

I want to evaluate two open source projects for this step: Hermes Agent and OpenClaw. Hermes Agent documents capabilities including VPS operation, persistent memory, skills, MCP integration, scheduled automations, and different communication channels. Source: Hermes Agent Documentation

OpenClaw pursues a similar always available approach through a gateway architecture. The gateway can run as a persistent service on a cloud server or VPS and connect models, tools, skills, plugins, and communication channels. Source: OpenClaw, Source: OpenClaw on a VPS

I have not yet chosen either system. My evaluation will focus on reliability, maintaining and backing up context and memory, interfaces, isolation, identities, secrets, logging, updates, backups, and explicit approvals for critical actions.

An always available agent with access to tools and interfaces is more productive than an isolated chat. It also increases the attack surface and the potential damage from an incorrect action. The goal is therefore not maximum autonomy, but a location independent workflow that remains available without sacrificing control and verifiability.

My current AI workspace is deliberately assembled. I adapted the LLM wiki, skills, integrations, and the roles of individual tools to my work step by step. Microsoft Copilot promises a different route: instead of assembling a harness from individual tools, it aims to provide one within a widely used work environment.

Microsoft describes Microsoft 365 Copilot as an AI tool that works in applications such as Word, Outlook, and Teams and can use work context that the individual user is permitted to access. Agents and connectors can additionally connect Copilot with knowledge, tools, and other systems. Source: Microsoft 365 Copilot overview, Source: Microsoft 365 Copilot extensibility

I therefore evaluate Copilot against the same five questions: Does it find relevant information in the work context? Does tool access reduce friction between question and result? Which tasks can it execute reliably? Can results be understood and improved? Are permissions, data sources, and responsibilities clear?

I am not starting this test by asking whether generative AI can be useful at all. I know the productivity of well designed agent harnesses from my daily work. I want to learn whether Microsoft can translate this quality into a standardized enterprise environment and which prerequisites must be met.

The Copilot test is therefore not an isolated product review. It is the next stage of the same investigation: what does an AI agent need to become more than an impressive chat in real work? My overview What is Microsoft Copilot? explains the product structure.

When adopting agentic AI, companies should not only ask which model leads a benchmark. The design of the complete work system matters more.

A productive agent needs a clearly scoped task, reliable context, appropriate access rights, controllable tools, and professional review. If any of these components is missing, either much of the potential remains unused or the automation creates new risks.

My own journey from GPT-3.5 through GitHub Copilot, Cursor, and Claude Code to Codex taught me one thing above all: experience with AI agents is not measured by the number of tools tried. It shows in the ability to turn models into repeatable, verifiable, and economically useful workflows.

Anyone who wants to assess how such an LLM wiki and agent harness could be applied to leadership, project, or knowledge work can use the free AI potential assessment to review a specific workflow with me.