Episode overview
In this episode, Ellis explores how AI agents, especially autonomous AI agents, are reshaping the landscape of technical communication. What are they? How do they differ from traditional AI tools? And crucially, what does their rise mean for technical writers?
Blending two recent blog posts, Ellis walks us through emerging tools like Manus, Opera’s browser AI, and AgentQL, and what these changes mean for how we create, structure, and deliver documentation.
What you’ll learn
- What AI agents and autonomous AI agents are — and how they’re evolving
- The four key traits of autonomous agents:
- Insights from new AI tools
Impact on Technical Writers
Ellis explores three main ways AI agents could change the role of technical authors:
- Content for autonomous AI consumption
- Structuring content for AI readability
- Multimodal delivery (text, audio, UI elements)
- Built-in accessibility for dynamic adaptation
- Documentation as agent-ready data
- Writing docs as if they were APIs
- Emphasis on semantic structure, metadata, and clarity
- AI agents as co-creators
- Personalised content generation
- Agent-driven feedback loops
- Enhanced content curation and adaptation tools
Key quotes
“Autonomous is the key word. It signifies something that’s self-governing, that can operate independently, with little human oversight.”
“The real transformative potential of AI lies in autonomous AI agents.”
“Rather than being replaced, technical writers could become the architects of AI understanding.”
Resources and links
Blog post: The Rise of Autonomous AI Agents
Blog post: Meet Your Future Co-Worker – AI Agents
Opera’s AI-Powered Browser Agent
The Race Is On to Redesign Everything for AI Agents
Transcript
We recently posted two blog posts on our website about AI agents. One was called “The Rise of Autonomous AI agents: will they redefine technical writing?” and the other one was called “Meet your future co-worker, understanding the rise of AI agents.” And in this episode, what we have done is we’ve combined those two articles into one, and what I’m going to do is talk about what they are, the new announcements that have come out recently, and the implications for what that might mean for technical writing. So let’s begin. IT is constantly changing, but one of the most important transformations isn’t just the word concept of AI or artificial intelligence; it’s the emergence, and the rapid emergence I should say, of AI agents and autonomous AI agents. AI agents could revolutionise how people interact with technology and the world around us, and they promise to become tools in our personal and professional lives. And autonomous AI agents promise to take that even further. The question for us as technical authors or technical writers is, what does this mean for the future of our profession?
You’re probably already interacting with AI agents.
For example, AI agents cover things like chatbots, which understand or perceive your question that you type into the chatbot. They decide on the relevant responses based on their training data and they act by displaying text. However, their autonomy is typically limited to what’s being programmed into them, into their pre-programmed conversational flows.
You may have also experienced recommendation systems on sites like Netflix and Amazon, where these services understand your viewing history and what you’ve bought in the past. And they decide which movies or which products you might like using complex algorithms, and they act by giving you recommendations. And again, for those AI agents, their autonomy is constrained, again by their algorithms and the available data.
Another area where you might be using AI agents today is if you use a grammar or spell checker, and these AI agents understand or perceive your written text, and they decide if errors exist based on the linguistic rules and models that they’ve been trained on, and they act by highlighting stakes and suggesting corrections. These types of agents have some independence to act and, as it were, think for themselves, but the actions are constrained.
Another area where we’ve seen AI used in technical communication is again to do constrained things like summarise transcripts, convert content from one format to another, describe an image or describe a video, convert a narrative piece of text into a set of step-by-step instructions, and also to create definitions for buzzwords, abbreviations, and other types of conceptual information. And even if AI doesn’t progress from this level, it’s still going to be a tool that is today and will be very useful for technical writers, in many ways, in a similar way to how Excel became useful for accountants and financial planners.
At this level, AI is useful, but it’s not transformative.
The real transformative potential of AI lies in what are called autonomous AI agents, and “autonomous” is the key word here. It signifies something that’s self-governing, that can operate independently, can do things with little human oversight. And that independence can happen across several different dimensions. So let’s go through those. One is evolving goals and sub-goals. An autonomous AI agent can break down a broad objective into smaller, manageable sub-goals. It can even redefine or adjust those goals as it learns and interacts with its environment.
Second aspect is that autonomous AI agents have more advanced decision-making capabilities. They can navigate ambiguous situations, adapt to unexpected events, and learn from past experiences to continuously improve their choices over time.
A third way is they can formulate complex action plans to achieve the goal that you set them, rather than just simply reacting to the triggers that they’re provided with. So they can strategise, they can anticipate challenges, and carry out multi-step plans without the need for constant human wanting. And as a consequence of that, be capable of longer-term thinking.
And the fourth way, which reinforces the other three that we’ve discussed, autonomous AI agents are also built to learn and adapt. They constantly analyze feedback from their environment and their own performance. They use this data to refine their strategies and improve their effectiveness. So this learning is continuous, it’s ongoing, it’s dynamic, and that means they’re able to be more proficient over time.
So these distinctions translate into important features that make autonomous AI agents different from the AI we’ve seen to date. So they have self-direction, they require less explicit step-by-step instructions, they’re able to figure out how to achieve a goal, yet just not what to do at each moment.
They have long-term vision. They’re capable of formulating and executing plans that span extended periods and involve numerous interconnected steps. So again, they move beyond immediate reactive responses.
They have better resilience and adaptability. They can effectively handle unexpected situations and environmental changes without needing constant reprogramming for every new scenario. So they’re designed to be robust and flexible in the face of changing circumstances. And the fourth difference we’ve discussed already, continuous learning, they typically incorporate machine learning mechanisms that mean that they are able to improve continuously, and evolve and become more proficient over time through experience. I mentioned at the beginning that there have been some recent developments with new tools that have autonomous AI agent capabilities.
What I’d like to do is talk about some of those to make it a bit more concrete in terms of what these things can do for you.
One that has created a lot of buzz is Manus, M-A-N-U-S. And Manus is designed to be a general-purpose, versatile AI agent. The idea is that you provide it with a task and it autonomously creates a step-by-step plan to execute it.
For example, you could give it the task to proofread and format a documenting word, and it will intelligently break down that task, identify the necessary steps which might be format checking and correcting errors, and then carry out those tasks using a variety of different tools.
Let me play some audio from a demonstration that has been put together by the creators of Manus.
Hi, I’m Peak from Manus AI. For the past year, we’ve been quietly building what we believe is the next evolution in AI, and today we’re launching an early preview of Manus, the first general AI agent. This isn’t just another chatbot workflow. It’s a truly autonomous agent that bridges the gap between conception and execution, while other AI stops at generating ideas, Manus delivers results. We see it as the next paradigm of human-machine collaboration and potentially a glimpse into AGI. Now let me show you Manus managed in action across three completely different tasks.
Let’s start with an easy one. In this example, we’ll ask Manus to help screen resumes.
I’ve just set Manus a zip file containing ten resume documents. Since each meta session has its own computer, it worked like a human. First, I’m zipping the file, then browsing through each resume page by page and recording important information and documents.
Manus works asynchronously in the file. Means you can close your laptop anytime and Manus will notify you when everything is complete. First, you can also get Manus’s newest structures at any time. Sheriff sent Manus five more resumes.
After carefully reading all fifteen resumes, Manus provides a ranking suggestion along with candidate profiles and evaluation criteria as supporting materials.
This is pretty good, but I prefer spreadsheets. Let’s have Manus greeting…
Manus has its own knowledge and memory, so it can teach Mattis that the next time they handle the similar task it will deliver a spreadsheet right away.
In this example, we’ll have Manus conduct some research. It needs to filter New York properties based on multiple criteria.
For complex tasks, Manus first writes down and creates a to-do list. Manus begins by searching and carefully reading articles about the safest neighbourhoods.
Then Manus researches middle schools in New York.
Next, Manus writes a Python program to calculate my budget.
Based on my budget, Manus filters listings on real estate websites.
Finally, combining all the information gathered, Manus writes a detailed report and compiles all the resources.
In this example, we’ll have Manus perform a correlation analysis between Stunks.
For professional data managed, Manus can access authoritative data sources at three PMs.
After validating required data managed, Manus begins writing code for data analysis and visualisation.
For Manus, coding isn’t necessarily a goal, but rather a universal tool for solving problems.
It looks like Manus has completed the data analysis and visualisation. But interactive data visualisation is even cooler. So I asked Manus to create a website based on these data.
With my permission, Manus deploys the finished website online and provides me with a shareable link. Let’s see what Manus has created.
And I’ll stop it there.
I should point out, at the moment Manus is only accessible by those that have an invitation code, and when it’s released for public access, it won’t be free. The pricing will be around thirty-nine dollars per month. The second announcement that caught our eye was from Opera, the company that makes a web browser. And in a press release, they said Opera has pioneered AI-powered agentic browsing with Browser Operator. They say this native AI agent can understand natural language requests and perform browsing tasks for users. For example, you could ask it to buy me a pack of ten pairs of white tennis socks from Nike, size twelve.
Browser Operator will then autonomously execute this task, navigating websites, searching for products, adding to cart, and proceeding to purchase.
And the third tool or announcement that caught her eye was from AgentQL. And Rachel Lee Nabors of AgentQL has, in a number of recent presentations at conferences, proposed a future where the idea of using a browser like Chrome will become outdated, and they see a future where personalised intelligent agents become the primary interface to the web. So let me play a segment from one of their recent conference presentations.
The internet. It started as a network of computers you could use to keep in touch with family, make new friends, chat, play games, share photos. I remember it as a friendly web, not a disorganised or a violent web, but then something happened. Platforms happened, Facebook, Instagram, Amazon. People moved off their blogs and forums and onto platforms where they were held hostage. So platforms, they promised ease of use. They promised great content, but it turns out that we went from free and open to gated and monetised. To keep users captive, they purposely lack interoperability, holding content hostage and people hostage with pricier no APIs and obfuscating data hiding it behind authentication. So yeah, now we all live in a bunch of gardens. So this means that to get anything done, this is actually one of my workflows which has gotten worse and worse over the years. We have to flip between a bajillion different platforms.
This is me just trying to keep up with web agent news from the agentic community. Mod a copy and paste in data entry. And if we want an aggregated experience, a one-stop shop, either we have to get other users and suppliers onto the platform we’re using, or we have to pay yet another platform like Buffer to give us that aggregated experience.
The internet is a vast sea of unstructured content with spotty and clunky interop at best. So why are we still using a browser built by an advertising company to navigate it?
Rachel Lee is suggesting organisations will use tool suites like AgentQL to connect their AI systems to the web, and this will enable the AI agents to do things like directly access and aggregate information from websites, and therefore bypass the platform interface and the adverts to get a more direct and more efficient experience, to be able to treat any page as an API endpoint, effectively transform websites into APIs. And this will allow AI agents to extract and interact with data programmatically, even if the website structure was intended for a more unstructured purpose.
It will enable us to personalise and curate content that we can use individual algorithms to filter and present information tailored to specific needs.
It will enable us to enhance accessibility dynamically. We’ll be able to, or using an AI agent, reformat content on the fly, to become audio, to become text, or adaptive UI, into a format that suits us as a user and our abilities.
Now this has many interesting implications.
How will walled gardens like Facebook or even direct airline travel sites like Ryanair react to AI agents getting over their walled garden and accessing their information? We may have a battle between those sites trying to stop autonomous AI agents getting in and breaking down their walls and making their information more accessible.
You might be asking what happens with privacy, with usernames and passwords? And there are emerging standards like the Model Context Protocol and the A2A protocol that have been designed to facilitate agent-driven web interactions, and to end up with decentralised social feeds. The model context protocol, or MCP, is an open standard developed by Anthropic, and it allows AI applications to connect with external tools, data sources, and systems.
It’s intended to make it easier to build AI-powered applications, like chatbots, assistance, and custom AI agents, the goal to creating a more open interoperable web. So MCP was designed to solve the problem, among others, of accessing private data on behalf of the user. It works by giving the AI tool some credentials, for example, an access token, to access the information on your behalf.
It has three main components that work together to enable this interaction to happen between the AI and the external system.
There’s a host. That’s the AI application or assistant that wants to interact with a website, your knowledge base, other data sources.
There’s a client, and this is the integral part of the host that understands and speaks the MCP language, and it manages the connection and the translation of the data between the two systems.
And the third bit is the server. That’s the system, the website that’s being accessed. That would be our knowledge base, a calendar, a database, and so on. So the host asks the question, posts the query.
The client translates that into a compatible format for the server and provides the authentication information, and the server delivers the relevant information.
So as I said, this could mean we end up going back towards a more user-centric, less platform dominated, and more accessible internet. And it could mean we give more control back to users to receive information and filter it and view it how they prefer.
But what about the role of the technical author, the technical writer? How could AI agents or autonomous AI agents reshape the work that technical writers do? Let’s look at that.
Number one, content for autonomous AI consumption.
As we’ve mentioned, the content we create is increasingly likely to be consumed not by human users alone, but also by autonomous AI agents operating on their behalf. And this could require a shift in how we approach documentation and focus on how we can make our content as it were AI readable.
How might that happen? Well, structured content could be one way. Although AI agents can process and manage messy unstructured content far better than traditional software, they do perform better when they have well-structured, semantically rich content to work on.
In the same way that some technical writers today have to consider SEO implications as well as end users, it might be that you need to consider creating content that’s intelligible to AI. If most of your users end up consuming the content you write via a chatbot, it might mean prioritising content for an AI system.
In practice, that means using headings and using metadata, and it also means using structured formats. And that could be simple ones like Markdown, and AsciiDoc, it could mean more semantic ones like DITA or ReStructuredText.
Another aspect is documentation beyond the visual. Considering multimodal delivery, we might need to think beyond traditional contents that’s primarily designed for on-screen reading. You might need to consider how our content can be consumed seamlessly in different formats, modalities, as audio summaries, as concise text extracts, as adaptive UI components. And autonomous AI units can do this. They can provide this dynamic delivery. But again, it may mean that the content we create needs to be structured in a way to support events.
And the third aspect, accessibility, inherent adaptability. Your documentation must be inherently accessible. So AI agents can adapt it for diverse user needs and abilities.
This is leading us towards the idea of the personalised, individualised help system where the AI creates a filter, presents the information, adapts the information specific to each individual’s preferences, abilities, and needs. So we need to consider, does the underlying structure and semantics allow AI agents to do this? Can they dynamically reformat? Can they repurpose content for various contexts and user needs? And another aspect to consider on this is something that’s been suggested by Tina He.
She is an entrepreneur at a company called Base, and she’s written a very interesting blog post on the every website called “The Race is on to redesign everything for AI agents.” And in that post or article, she asks, what happens when agents supplant humans as your primary users and customers?
She puts forward the idea of AX, or Agent eXperience, and suggests great AX is when an agent performs a task exactly as you wanted it to, and it can perform everything it needs, to the first time it’s asked. And it can do this in a way that’s cost-effective without the need for human intervention.
And she argues that achieving that goal takes careful consideration of different criteria such as onboarding. Agent onboarding involves verifying permissions, providing secure access tokens, and offering structured documentation that AI can interpret, developer kits. When building a software development kit for humans, you focus on intuitive APIs, detailed error messages, and comprehensive examples that mirror real-world use cases.
Agents, she argues, however, need standardised machine-readable product descriptions, explicit instruction flows, and robust metadata so they can understand and take advantage of your tool’s functionality.
And the third criteria she mentions relates to interactions and permissions. She writes you need to make sure that when an agent connects to your system, it can prove it’s a good actor, and that everything it does can be audited in case something goes wrong.
So, those criteria, those aspects suggest we’re going to, or organisations are going to, need more content, and someone is going to need to write it and be able to write accurately and to the quality that’s required.
Number two, documentation as an agent-ready data source, thinking API, not just web page. So this means structuring for what you might call plausibility. Think of your documentation not just as a set of web pages, but as a data source that AI agents can query and extract information from. So that could mean structuring it in ways that are possible, inherently possible, and easily understood by AI agents.
It might mean more granular content components and again, more semantic markup.
Comprehensive metadata could become more crucial to enable the agents to discover and understand the purpose and content of your documentation. And another aspect is accuracy and conciseness. Agents rely heavily on the accuracy of your documentation to provide reliable information to users. Conciseness and clarity become critical, as agents might summarise and present information in condensed forms, and if there is ambiguity, that can lead to misrepresentations, misinterpretations by the AI agents and the users that they’re serving.
We’ve looked at two possible changes, that we might need to write content for machines for autonomous consumption, and that we might need to think about our documentation as if it were an API because it might be consumed in a way similar to how APIs are used to extract information. The third aspect or third possible way in which AI agents and autonomous AI agents could affect the technical writing role is as a co-creator in the writing process itself. So beyond creating content for agents, technical authors, technical writers can use AI agents to enhance their own workflows and processes.
Now, the fear is, for many technical writers, is that AI will make their job redundant, that it will be possible, or it might be possible to tell an AI system to make a help system, and whoosh, it’s there. So will autonomous AI agents make that concern real?
It’s two aspects to consider there. One is personalisation and contextualisation. AI agents could dynamically deliver and adapt documentation based on individual user context and needs.
As we’ve mentioned, it could make documentation more personalised, relevant, and effective more than ever before. And that’s a good thing.
If we imagine documentation that adapts in real-time to users’ roles and experience levels or specific tasks, it makes it more effective documentation.
Now, if you’re writing multiple documents for different audiences, then the AI system may take that role away from you, but for many organisations, they have to have a one-size-fits-all manual for everyone. And by using autonomous AI agents, they can get around that limitation, that challenge of limited budget and time, and create individual content, individual experiences for each user.
Another aspect is agent-driven feedback. Agents can analyze user interactions with documentation and with the UI, and provide data feedback on areas where the UI or the documentation is unclear and might need refinement. So AI could help provide the ability to have continuous improvement and optimisation of your content.
What this suggests is that while some tasks within technical writing might become automated by AI, the core skills in communication, information architecture, and understanding user needs will become arguably even more valuable if we want an AI agent and an autonomous AI agent to work effectively and to the best that they can be.
Autonomous AI systems still need guidance. They require clear instructions, definitions, and conceptual frameworks to make the best choices and deliver the best useful information. Again, they need accurate information. So there’s going to be a need for what you might call architects of AI understanding. And it could be that those architects are technical writers. The role might develop and evolve to include things like developing precise and effective prompts to guide the AI to create content, to create user assistance. And there will be a need to review and refine AI generator content to make sure it is accurate, it’s consistent, and it’s in alignment with what the users want, that it’s consistent with good communication principles. And again, we have this issue that agents must have access to accurate and well-structured knowledge to become effective.
It seems like the developments in AI agents and autonomous AI agents are likely to be a fundamental shift in how we as users interact with technology and information, and also how information itself is created and consumed. And for technical writers, it might feel like it’s a predator or prey situation. Are we the predator, or are we the prey? But in the future, we might look back at this time as a massive opportunity for professional growth and evolution rather than obsolescence.
To wrap things up as a summary, what can we say?
The future of the web, the internet, and of software interaction is going to become increasingly intelligent, proactive, and driven by AI agents. And people who have skills in communication and information design are well-positioned to be at the forefront of this change.
By embracing content, well-written structured content, by prioritising accessibility, and expanding your skill set to include AI-assisted tools and workflows and understanding the principles of AI-driven, AI agent-driven content delivery, then there’s an opportunity to not only adapt but also thrive.
There’s some key words in there, and that is to be able to adapt, to learn, and to position yourself as the expert, the essential guide in this new world to shape the capabilities of the AI system and to define how the information is created and consumed, both for the end users, the humans, and the intelligent agents that they and us are going to be working alongside.
If you have any comments on this, then let us know. You can email us at info at cherryleaf dot com. We’ll provide links in the show notes to the Opera browser, to Rachel Lee Nabors’ presentation, to the Manus website, and also to the two blog posts that we’ve written.
If you’d like to know more about Cherryleaf and our technical writing services and some of the AI-related stuff that we do, you’ll find that on our website, Cherryleaf dot com.
And…
Leave a Reply