Individuals Including AI and Interactions Over Processes and Tools

Agentic AI Developer models like Claude and Gemini are trained on vast quantities of data and can be thought of as repositories of knowledge accessible through a conversational interface. When you start a new conversation all of that knowledge is available to you, but the model comes with no understanding of your collective context to apply when formulating its responses.

If you are asking the model for the definition of a word or to tab-complete a line of code, a low or no context state has little impact on the quality of the response. You are essentially using the model as a tool: there is little need for nuance. A hammer rarely needs to understand much about the nail.

On the other hand, if you are asking a model to code a front end in React for a web application, then a contextual understanding of the web application becomes much more important. What is the value proposition of the application? What are the functions that need to be surfaced? What is the persona of the target user? All of these questions inform the implementation. A developer needs to understand many contextual details to effectively contribute to a project, and this is equally true whether the contributor is a human developer, or an Agentic AI developer.

A way to think about the various models of interaction with Agentic AI development technology is to recognize that the more autonomy we are seeking for Agentic AI developers to display, the more we wish to anthropomorphize them and interact with them like they are human, the greater the requirement is that we provide comprehensive context to the Agentic AI developer, not just about the task at hand, but about the situation it is in.

The Agile Manifesto starts with the words:

"We are uncovering better ways of developing software by doing it and helping others do it. Through this work we have come to value: Individuals and interactions over processes and tools..."

The act of coding is a deeply human activity, full of discovery and learning and creative energy that no amount of planning can harness and bend to an arbitrary schedule. The Agile Manifesto was a courageous declaration that developers are in fact humans, who work best when levering human behaviors and acknowledging human needs.

The accompanying Principles behind the Agile Manifesto have more to say about this. For instance, the principles include:

Business people and developers must work
together daily throughout the project.

Build projects around motivated individuals.
Give them the environment and support they need,
and trust them to get the job done.

The most efficient and effective method of
conveying information to and within a development
team is face-to-face conversation.

The best architectures, requirements, and designs
emerge from self-organizing teams.

The emphasis on face to face communication on a daily basis between motivated members of a trusted self-organizing team is core to the principles of Agile.

So what now, when ever greater numbers of developers are discovering the possibilities inherent in using AI development tools to help 'satisfy the customer
through early and continuous delivery of valuable software?' Does the use of these Agentic AI development tools eliminate the need for business people to have to talk to Engineers every single day? Can we also get rid of the need for all that uncomfortable face to face sharing? Can we finally now stop having to trust teams, since we can just tell a machine what we want the software to do?

A big misconception about software development common among people who do not actually do it is that the work of software development only occurs when hands are on keyboards writing code. This view of software development sees meetings as a waste of time, because meetings are a means of driving alignment between people about what that software is going to do, and the people should already somehow know that. In this mental model of work the keystrokes are all that matter.

In fairness, as the diagram below shows, processes like scrum do incur some meetings overhead in order to maintain an orderly flow of information through the process.

These process oriented meetings are in addition to the myriad of meetings required to actually perform the work of development, like meetings between developers working on related items of code, or with product managers to ensure the functionality is properly understood and instantiated.

Often the question arises, if we are already having all of these working meetings, why do we need to also have these process-oriented meetings? There are too many meetings already, and surely with that many meetings any information needed will filter down to the people who need it...

This is a shorthanded way of saying that establishing and maintaining the context of the work - making sure we all understand and agree on who is doing what - to accomplish which objectives - and why - is less important than the work involving actual typing.

Or put another even shorter way, we want to consider you the developer as more like a hammer than a human, and do not think it is efficient to allocate any time or effort for you to understand much if anything about the nail.

Hybrid Agentic AI models are where you provide an AI developer Agent to a human software developer and then expect them to proceed to write code like two humans would, in a sort of paired programmer model.

The expectation that this will be intuitively successful belies a fundamental misunderstanding of the situation which has been created.

Humans are able with effort to develop a persistent and evolving sense of personal context over time. Through social interaction, research, and experimentation, humans discover the world around them and apply their knowledge about it in their day to day activities.

When people say that it is inefficient to devote time or resources to the development and maintenance of context for human workers, it is because they expect those workers, due to their fundamental nature as humans, to already be gathering and processing this contextual information on their own.

AI Development Agents throw a wrench into this conception. Every time an AI Development Agent is activated, it arrives with no contextual understanding of any kind. The AI development agent can only ever establish any sort of useful context by reading documents containing contextual information. These documents, or instructions to find them, must be explicitly provided.

Context for AI Development Agents is limited in size. The maximum size of the context is referred to as the 'Context Bucket' and is measured in tokens. A token is a computed value, not exactly a number, not exactly a word, which for the AI Agent is translatable to some semantic relevance when read along with lots and lots of other tokens. A typical context bucket today is somewhere between 50k and 1M tokens.

Every interaction with the AI Development Agent contributes to the contents of the context bucket. Over time the original contents of the context bucket become diluted, as newer interactions replace contextual data already in the bucket from previous interactions. Practically, this means that if you establish the context for an item of work - the functional requirements, the target persona, the why - as that item is being worked on by an Agentic AI developer its ability to recall that context becomes irreversibly lessened over time, until it is altogether lost.

When the context becomes so diluted that the Agentic AI developer can no longer remember why it is doing what it is doing, it becomes necessary to establish context once again, ideally by feeding it the same documentation in the same order using a repeatable process, so that the consistent establishment of context produces consistent contextual understanding as a result.

The more that relevant context is in place, the more the Agentic AI developer is capable of acting with nuance, offering relevant suggestions, and creating value independent of its human co-workers. The less that relevant context is in place, the more likely the suggestions and actions will be off the mark, at very least, or spawn chaos, as the agent proceeds to act in an undirected fashion with the benefit of knowledge alone.

So the precise thing we quite typically deny human workers - time and effort to develop and establish and maintain context - is utterly necessary if we want Agentic AI Development agents to do more than hammer nails.

This is where the true modern value of Agile process becomes evident. Repeated establishment of context is a requirement to enable the full benefit of Agentic AI developers in a paired programming model. To consistently accomplish this the contextual data has to come from somewhere, but where?

The obvious answer is the Agile process, which was originally designed to establish and maintain precisely the same context, but for humans.

As we have seen it is easy to diminish the context-establishing value of Agile process by belittling the need to support human workers in this basic need. With Agentic AI workers no similar statement of avoidance is possible: Agentic AI workers arrive, at every working session, with no contextual understanding, and are incapable of maintaining for any significant length of time the contextual understanding they are proactively provided.

Optimizing for unexpected and positive contributions by Agentic AI contributors requires optimizing for the depth and quality of the contextual information they are repeatedly and consistently provided. Contextual information which can only result from consistent and proactive human participation in a software development life cycle methodology like Agile or Scrum or Scaled Agile or... take your pick, the actual flavor probably doesn't matter as much to this outcome as the requirement that you pick one, and you use it to generate the contextual information required to inform Agentic AI developer co-workers about the details of their current, ephemeral, instantiation.

The Agile Manifesto says

The most efficient and effective method of
conveying information to and within a development
team is face-to-face conversation...

Adding Agentic AI developers to the mix adds a simple requirement which makes the statement a whole lot less pithy:

... the transcript or results of which are written down so they can be given to Agentic AI developers who could not participate in the conversation, so they also can benefit from the interaction.

To realize the full value of investments in Agentic AI developer technologies we're going to have to do things like document meetings and proactively maintain records of project and program state, which we are pretty consistently reluctant to do for projects which involve human workers alone. Fortunately the Agentic AI developers are extremely fast and efficient at this sort of documentation, reducing the burden of this requirement tremendously.

The irony of this is that, if we do ask the Agentic AI Developers to consistently maintain the written records needed to maintain their contextual state, using information from Agile process meetings necessary to develop, document, and maintain that contextual state, the human workers they work alongside will benefit enormously. The consistent and more transparent flow of relevant information will enable the human workers on the team to also contribute in ways which more closely reveal their own inherent human capabilities.

Most AI projects fail—not because of the technology, but because the organization wasn’t ready. We help organizations align people, processes, data, and technology for long-term success in the intelligence era. Our free assessment will help you evaluate your company’s readiness to implement AI in a scalable, responsible, and impactful way. Click the link, or give us a call, and lets move forward together to realize the opportunity AI holds before us.