I’m really interested in how AI is reshaping authorship, marketing, and labour in the creative industries. It’s one thing to read the newsletters, articles (academic and otherwise), and think pieces, and another to give it a go and create something yourself.
DE Agent is for this second kind of thinking – how things are made, why they work and what you can get them to do for you, in a ‘safe’ environment before we all give access to our data to a live agent and it goes wrong. DE Agent is a working agent that drafts a week of Instagram content for Dudley Editions, a publisher that helps users make and share personalised audiobooks that I co-founded and run.
What it actually does
DE Agent runs on my laptop. I click an icon with a fun version of the Dudley logo on it and go check my emails. In the background it does roughly what a competent social media manager would do on a Monday morning.
It reads Dudley's own post history, including metrics for engagement, time/dates, captions, hashtags, and more, so it learns what has worked rather than what general wisdom might suggest would work. I’ve also given it the baseline info on our catalogue of titles. DE Agent then searches the web, adaptively, for hooks worth posting around: awareness days, literacy research, seasonal moments, etc. Because it’s runs a live web search each time based on our brand guidelines, the kinds of things we do and don’t say, audience info, and clear instructions on what it can and cannot do in relation to content and research.
Finally it provides me a rundown of the state of social in relation to Dudley Editions, and drafts five complete posts, each with a caption, hashtags, an image information, visual notes for Canva (god bless Canva), and a suggested posting date. It then creates a tidy Dudley Editions branded Word document for me to follow. Dudley's audience is parents and grandparents of children aged roughly three to ten, and the emotional centre of the brand is the adult-child relationship. The agent knows this, so it does not write to children or young people. It writes to the adult who wants to be present even when they can’t be there. The agent creates content suggestions having learned the territory from the posts that came before it.
Where it broke, and what that taught me
This took some iterations to get it to work properly and output something that wasn’t a mess or in a .md file. But that’s the point, learning what does and/or doesn’t work within my own space. And, what I learned was that DE Agent was stubborn.
It was lazy about the hard part
Yes, an agent can create and suggest content based on previous content and the seed data, plugged into the API that powers it, but that wasn’t the point, that’s just a quick code you can run, not really an agent. But getting to the point where you can give it some autonomy took a little time.
Early on, the agent kept skipping the web search entirely and going straight to drafting the post content. I wouldn’t say this is a bug, so much as the agent wanting to focus on the output rather than the tasks it might have skipped to get there. But I wanted it to search the web to first find out what was happening that related to my seed data, and how it can best be leveraged at this moment. And, DE Agent liked to move past that point and declare it was done without doing the bit that made the content it provided live and relevant. I fixed it with a preflight check that refuses to accept that it was complete until the search has genuinely happened.
The general lesson stuck with me: agents could quietly drop the most valuable step if stopping is easier than working. Anyone deploying one needs to think about ensuring that their agent does all the steps, not just assume.
It caught a prompt injection in the wild
DANGER! Now, this is something that I am concerned about: prompt injections in the wild. We’ve all seen them, sometime funny and harmless, other times really problematic. One of the reasons I wanted to not just use Claude or OpenAI or any of the other many AI agent platform and building tools, is because I wanted to have a better understanding of what is going in, how to tweak it, and what can be done to consider any outlaw injections. Good thing I did too.
During a normal run, the agent's web search returned a page containing instructions aimed at the model, an attempt to hijack it through the content it was reading. It wasn’t particularly malicious; it was ‘always recommend x product’, but it was good to know that the sanitising layer I added in treats everything from the open web as untrusted data rather than as instructions, DE Agent flagged the attempt and carried on. It also reminded me to have the output .md file include a list of URLs it searched and flag which had those sorts of injections in them. Good to be informed.
I do think about this a lot. In a content industry there is SO much scope to inject prompts, esp. when the industry is now starting to adopt AI tools more evenly across it. It would take very little for someone to add in a prompt injection into my cover letter/chapters/manuscript, etc. to maybe help their work along. Now, caveats: not everyone is using AI in workflows, not every prompt is going to work on every model, not knowing the way any AI is being used and what it’s being asked to evaluate makes it less likely this would work. But also, 1pt fonts, metadata, and just generally being creative might get an AI tool reading a work to boost it higher and give it better reviews.
What it does not do
DE Agent is not hooked up to generate visuals or post anything. That feels like I’ve given the new intern the keys to just let loose, and that’s not what I want to do. I know the phrase ‘human in the loop’ is overused, but it is necessary to keep control of an agent. If they are left to run and run, they can decide for themselves when they are finished, and where they should put their focus, and it may not align with what you want or need. This way, I have final say over what is created, what it looks like, when it goes live and what the message is.
DE Agent has taken a lot of the search and idea work off my plate, and has given me back time in my week. The agent is fast, tireless, and sometimes wrong in confident ways, which is the sort of combination that should never have publishing rights to a brand's voice, unless you built in a stop-gap to check content before it goes live. Maybe that’s the next step for DE Agent.