What I mean when I say I work agentically

I have been working agentically for several months.

That sentence probably needs explaining.

“Agentic” is rapidly becoming one of those words that makes sense to people working closely with AI and means almost nothing to everyone else. It can sound like a more impressive way of saying you use ChatGPT.

That is not what I mean.

I use ChatGPT, Claude Code and GitHub. But the agentic part is not any one of those tools.

It is the way I have connected them into a working system that can remember decisions, develop ideas, make changes and move work forward.

I am still responsible for the thinking. I decide what matters, what I believe and what gets published.

But I am no longer personally carrying out every step between having an idea and turning it into something useful.

What does agentic AI actually mean?

A conventional AI assistant responds to a request.

You ask it to summarise a document, draft an email or suggest some ideas, and it returns an answer.

Agentic AI can go further. It can work towards a goal through several steps, use tools, inspect information, make decisions within defined boundaries and take actions in other systems.

That might mean searching a repository, finding the relevant file, editing it, checking the result and opening a pull request for review.

An agentic workflow connects those capabilities into a repeatable way of working.

It does not necessarily mean giving AI complete autonomy. In my case, the agents can investigate, propose and implement changes, but I still set the direction and approve what becomes part of the live system.

The simplest distinction is this:

An AI assistant gives you an answer.

An AI agent can help move the work itself forward.

A practical example

Editorial Intelligence began as a collection of ideas about research, editorial judgement and organisational knowledge.

It could easily have remained a collection of documents.

Instead, I created two GitHub repositories.

One holds the private methodology: frameworks, research questions, working principles, content ideas, observations and decisions.

The other holds the public website.

ChatGPT works with the accumulated context around the project. I use it to develop observations, challenge arguments, identify connections and turn rough thinking into possible articles, frameworks or changes to the site.

When I decide something should be recorded or changed, I can ask ChatGPT to work directly with GitHub.

It can inspect the relevant repository, find the right file, make a proposed change and open a pull request.

Claude Code gives me another way to work across the same system. It can inspect the repository, challenge the structure, edit the writing, test the website and look at the implementation from a different angle.

Then I review the result.

Sometimes I publish it. Sometimes I rewrite it. Sometimes I decide the original idea was wrong.

The workflow looks like this:

1. I notice something.
2. ChatGPT helps me develop and connect it.
3. GitHub gives the thinking somewhere persistent to live.
4. An agent changes the relevant part of the system.
5. Claude Code provides another editorial or technical pass.
6. I make the final judgement.

This is not a fully autonomous machine producing content while I do something else.

It is a working environment in which different agents have context, roles and access to move a project forward.

Why this is different from using an AI assistant

Most everyday AI use still looks like a transaction.

You provide a prompt. The model gives you an answer. You copy the answer into another tool. The conversation ends.

The next time you return, much of the important context has to be supplied again.

That can make AI useful without changing the underlying operating model. You are still coordinating everything manually. The AI helps with individual tasks, but the human remains the connection between every system.

An agentic workflow changes that relationship.

The model is not only generating an answer. It can inspect the environment in which the work lives, understand the current state, take an appropriate action and leave the system in a more developed state than it found it.

The output might be an article draft. But it might also be:

• a research question added to a backlog;
• a framework updated to reflect new evidence;
• a website page reorganised;
• a broken link identified;
• a pull request created;
• a previous argument connected to a new observation;
• or a decision recorded for another tool to use later.

The difference is continuity.

The work does not disappear when the conversation ends.

GitHub became more than a code repository

GitHub is usually explained as somewhere developers store and manage code.

I increasingly use it as somewhere thinking can accumulate.

It holds the website, but it also holds the decisions behind the website. It records how the methodology is changing, which ideas need more evidence, what I intend to publish and what I have learned from the response.

That persistence matters because language models are good at synthesis but need material to synthesise.

Without a durable evidence base, an AI system is forced to work from whatever happens to fit inside the current prompt. It can sound intelligent while repeatedly starting from the beginning.

With a maintained repository, old thinking can participate in new thinking.

An observation about a LinkedIn post can connect to an earlier idea about editorial judgement. A customer interview can change a framework. A new article can reference an existing argument without me having to remember every connection manually.

The repository is not simply an archive.

It is part of the workflow.

Why I use more than one model

I pay for both ChatGPT and Claude.

That may look inefficient because their capabilities overlap. But I do not treat them as interchangeable writing machines.

ChatGPT has become the place where I develop ideas over time. It knows the wider context, notices connections and helps translate incomplete thoughts into something I can examine.

Claude Code works directly across the project and often approaches the material more like a second editor and implementer. It is useful for reviewing structure, challenging a page, changing the site and checking whether an idea has been implemented properly.

This gives the work three editors and four passes, rather than asking one model to draft and approve its own output.

The models will change. Their relative strengths will change too.

That is why the workflow should outlive the model.

The durable layer is the evidence, decisions and working structure stored outside them.

The agents can change. The system should survive.

This is still an editorial process

There is a temptation to describe agentic work as removing the human from the loop.

That is not what makes it valuable to me.

The value comes from being more deliberate about where human judgement belongs.

I do not need to manually move every paragraph between tools. I do need to decide whether the paragraph says anything worth publishing.

I do not need to remember every relevant document. I do need to judge whether the connection an agent has found is meaningful.

I do not need to make every website change myself. I do need to decide what the website is for.

My process currently has four editorial passes:

1. Me, originating the observation and deciding whether it matters.
2. ChatGPT, developing the argument using the accumulated context.
3. Claude Code, reviewing or implementing it from another perspective.
4. Me again, deciding whether I can stand behind the result.

AI has not removed editorial judgement from this process.

It has made the location of that judgement much clearer.

Why brand teams should care

Brand work depends on consistency, but consistency is often maintained through documents that people have to remember to consult.

There may be positioning decks, tone-of-voice guides, messaging frameworks, customer research and campaign strategies. They exist, but they do not necessarily participate in the work.

An agentic brand system could make those materials active.

An agent could compare a proposed campaign with the agreed positioning. It could identify when claims lack supporting evidence. It could find relevant customer language, flag contradictions and record decisions so the next piece of work starts from the updated position.

That does not mean automating brand judgement.

It means giving brand specialists a system that can carry more of their judgement into everyday execution.

The brand stops being a document people occasionally remember.

It becomes part of how work moves.

Why performance teams should care

Performance marketing already has data, automation and optimisation.

But optimisation can become detached from organisational learning.

A campaign performs. The numbers go into a report. A few changes are made. Then the next campaign begins with only a partial memory of what happened.

An agentic workflow could connect performance signals to the wider evidence base.

It could record which messages generated attention, distinguish a strong click-through rate from a meaningful commercial result, relate performance to audience research and surface patterns across campaigns.

The important change is not faster reporting.

It is turning campaign response into reusable knowledge.

That creates a stronger connection between performance and brand. Performance supplies evidence about what audiences respond to. Brand supplies judgement about what the organisation should say and what it should not become merely because a particular message produced clicks.

Neither side should operate without the other.

What agentic does not mean

There are still obvious weaknesses.

Agents can misunderstand instructions, make weak connections and confidently change the wrong thing.

Persistent context can also preserve bad assumptions. A system that remembers everything is not necessarily a system that understands anything.

This is one of the problems Editorial Intelligence is designed to address.

Agents need more than access to information. They need an evidence and judgement layer around that information:

• which sources should be trusted;
• which claims are supported;
• where evidence is incomplete or contradictory;
• which existing narratives and decisions matter;
• what an agent is allowed to change;
• where human review is required;
• and how mistakes are corrected in the underlying system.

Editorial Intelligence does not guarantee that an agent will be right. It makes the basis of its work more structured, visible and reviewable.

In my own workflow, GitHub provides the technical control layer. It records what changed, preserves previous versions and uses pull requests to create a moment of inspection before anything becomes part of the live system.

But GitHub cannot judge whether an argument is supported or whether a connection is meaningful.

That is the Editorial Intelligence layer: the accumulated evidence, frameworks, narrative logic, decision history and editorial criteria against which the change can be assessed.

A second model provides another perspective. I make the final decision. If something turns out to be wrong, the correction should change the stored knowledge rather than simply repair the published output.

GitHub shows what changed.

Editorial Intelligence helps decide whether that change deserves to survive.

This is slower than pretending an agent can run everything alone.

It is also how an agentic workflow becomes a learning system rather than an automated production line.

The real change is the operating model

The most interesting part of working agentically is not that I can produce articles faster.

I can, but faster production is the least consequential result.

The bigger change is that one person can operate a connected system of research, thinking, publishing, development and review without manually performing every part of the process.

That changes the unit of capability.

The useful question is no longer only:

What tasks can AI complete?

It is:

What kind of working system can we create when people, agents, evidence and tools can act on the same accumulated context?

I built mine around a small editorial project.

The same principle could apply to a brand team, a performance function or a larger content operation.

But it starts with something less exciting than autonomous agents.

It starts by giving the work somewhere to remember what it knows.