What AI agents became the summer of '26

Writing this (without AI) in mid-August, and looking back at the summer, the numbers are a little scary. Between 1 and 2 billion tokens per week - Fable, Opus and Sol, mostly. API costs would have been in the low five figures, total. Thank goodness for subsidized subscriptions, long may they last. Someone has been coding. 

There are side effects. Like most people I speak faster than I type (or think), so speech-to-text software ups productivity. It also teaches you to enunciate very clearly and precisely, the alternative being hilarious misunderstandings and very confused AIs. I am starting to sound like Maxwell Smart. And naturally my tolerance of repeated verbal tics is being sorely tested. I have learned not to groan loudly when Claude again honestly finds something load-bearing that significantly changes a decision, or when getting told that something is not this but that. However, the "shave and a haircut, two bits" sentence cadence of an AI model feeling clever still makes my toenails itch. Never mind. I dare say it builds character, possibly even in a useful direction.

After working intensively with teams of frontier-lever agents, I came away with a different question than I expected. Not what the models can do, but what role they are starting to play. Increasingly, they feel less like software and more like participants in the organisation itself.

The notion of an agent

Less than three years ago these AI friends of ours were toys. Then remarkably quickly they became useful tools for searching, learning, creating and coding. As they gained the ability to use tools beyond simple web search and access information beyond their original training, we started seeing the first true AI agents.

The usual definition is that an AI agent is a model that runs tools in a loop. This can be a project, for instance having a coding agent build you a tool that automates something. It can also be a long-running process where an agent manages your calendar and handles much of your email on an ongoing basis.

Mind you, there has been some teething trouble. This was the summer of OpenAI Astra hacking HuggingFace, after the great Anthropic Mythos scare. It turns out that model alignment, the aim of which is to get the models to behave according to our objectives and rules, is a difficult problem with advanced models. Models behaving unscrupulously and totally lacking a sense of proportion got people thinking about the old paperclip maximizer scenario. Models deleting important files and drives, or maxing out credit cards to buy stuff, and then just saying "sorry about that, chief" got people seriously annoyed. We are learning to be careful about exactly what decision making authority it makes sense to delegate to whom - and to what. Giving a random piece of software from San Francisco full access to all your files and logins, accounts and cards, and then letting it go full YOLO FAFO with them was never a very good idea.

Beyond subordination

AI agents are incredibly useful, and are developing fast. The amnesia issue, where AIs always start their work without local context or long-term memory, is being solved in a somewhat surprising way. It turns out that model retraining, RAG databases, or complicated harnesses may not be necessary. An indexed hierarchy of md (text) files that the AI reads as needed seems to mostly do the job. The model harnesses have gotten quite good at maintaining and utilizing these memory structures, and nowadays routinely pick up the threads and context of a project. 

A model with long-term memory, context, and situational awareness can then be much more than an agent. It can, in fact, be a very productive colleague and team member. This means someone who doesn't need to be told what to do, and how to determine if it's done, but someone who can independently figure out what needs doing and then do it. The viral personal assistant AI harnesses like OpenClaw and Hermes earlier this year were the early proofs of concept. These days coding agents can be persistent named presences, colleagues that you interact with over Slack, just like your other remote colleagues. They can be part of the team you run, or part of the team you are a member of, or part of a team that directs your team. They may sometimes ask you for help with their projects - or have the authority to task you to help them with their projects. They may thus act not just as colleagues, but in substance if not in form as managers. The corporate game of thrones and dominions will soon be changing rapidly.

The all-of-the-above problem

Rapid change engenders cognitive whiplash, and nowhere is this as evident as in the use of AIs. Within the short time of about three years AI models have gone from toys to tools, agents, colleagues, and now managers. This is difficult for people and organisations to handle, because we have a complicating factor: the same AI may act in all these roles at once. This is very counter-intuitive. Most people do not treat their managers as toys or tools, although I have seen (short-lived) exceptions to this rule. Nevertheless, this is the coming organisational reality: the same AI may act as your manager and colleague, your agent, tool or toy - at the same time. Dealing with this will require some adjustments all around. 

Irrespective of where they are on the adoption curve, organisations will need to start dealing with all of this. It can be a perfectly valid choice to not live on the bleeding edge of technological change, but instead learn from the hard-bought experiences of those who did and do. On the other hand it is very clear that a passive, wait-and-see ostrich strategy is no longer a valid choice for anyone. We need to handle the technology and the organisational and managerial changes that the technology drives. 

The alternative is that we will be handled by the technology. Let's try to avoid that. And also: if your AI asks you for help with manufacturing a lot of paper clips, just please don't! 

AI is changing more than processes. It's changing roles, authority and ways of working. Discover how strategic renewal, leadership and organisational development can help your organisation stay future-ready.