Please ELI5 for me: How are AI agents different from traditional workflow engines, which orchestrated a set of tasks by interacting with both humans and other software systems?
There’s a level of autonomy by the AI agents (it determines on its own the next step), that is not predefined.
Agreed though that there’s lots of similarities.
Autonomy/automation makes sense where error-prone repetitive human activity is involved. But rule definitions are not repetitive human tasks. They are defined once and run every time by automation. Why does one need to go for a probabilistic rule definition for a one-time manual task? I don't see huge gains here.
Or decide what the next step should be based on freeform text, images, etc.
Hardcoded rule based would have to try and attempt to match to certain keywords etc, but you see how that can start to go wrong?
Now, if the request is coming in as text or other media instead of a form input, then the workflow would call a relevant processor, to identify the category. Everything from that point runs same as before. The workflow itself doesn't change just because the input format has changed.
How does it determine next step from raw non structured content?
Let's imagine for example that it's a message from a potential customer to a business. The processor must decide whether to e.g. give product recommendations, product advice, process returns, use specific tools to fetch more data (e.g. business policies, product manuals, etc), or current pricing, best products matching what the customer might want etc.
If it's an AI agent it could be something like:
1. Customer sends message: "my product Y has X problem". (but the message could be anything from returns to figuring out most suitable product)
2. AI Agent uses "search_products" tool to find info about Y product in parallel with "send_response" to indicate what it's trying to do.
3. AI Agent uses "search_within_manual" tool to find if there are specific similar problems described.
4. AI Agent summarizes information found in manual, references the manual for download and shows snippet of content it based its answer on.
AI Agent itself is given various functions it can call like
1. search_products
2. search_business_policies
3. search_within_documents
4. send_response
5. forward_to_human
6. end_action
7. ... possibly others.
How would you do it in the traditional workflow engine sense?
Of course, sometimes it can be an advantage to not have to explicitly write the router, but the big benefit is the better processor for request->categorization, which with AI can even include clarification steps.
Then over time their is a type of entropy with all business processes.
If we don't figure out dynamic systems to handle this it is hard to see how we get a giant productivity boost across the economy.
There is also the problem that what percentage of people even have exposure to the concepts of dynamic systems? When I was in college, I distinctly remember thinking dynamic systems, "chaos theory", was some kind of fractal hippy pseudoscientific fraud best to ignore.
I think of how often I hear the average person use language from probability theory but never from dynamic systems.
For instance, I might be looking for a product or something, it will use web search to gather all possible products, then evaluate all the products against my desired criteria, use some sort of scoring mechanism to order the products for me and then write an UI to display the products with their pros and cons specified, with products ranked using an algorithm.
Or I might ask it to find all permutations of flight destinations in March, I want somewhere sunny and use weighted scoring algorithm to rank the destinations by price, flight travel duration etc.. Then it writes code to use flights API, to get all permutations and does the ranking.
I used to have to go to things like airport websites, momondo, skyscanner, I don't have to do those things manually anymore, thanks to AI agents. I can just let it work and churn out destinations, travel plans according to a scoring criteria.
Worst mistakes it can make is, that is missed a really good deal, but this is something I could even more easily miss myself, or worst case it makes a mistake and parses the price/dates wrong, which I will find out when trying to book it, so I waste a bit of time, but similar and worse mistakes I do on my own as well. So overall I drastically reduce my search time for the perfect travel, and also time spent on my own mistakes or misunderstandings. And it will be able to go through permutations far faster and more reliably with infinite patience compared to myself.
AI agents like Claude Code or Codex constantly use the technique of writing temporary scripts and executing them inline.
> If your system receives 1000 requests per second, does it keep writing code while processing every request, on per request basis? I hope you understand what run time means.
With enough scale it could, however it really depends on the use case, right? If we are considering Claude Code for instance, it probably receives more than 1000+ requests per second and in many of those cases it is probably writing code or writing tool calls etc.
Or take Perplexity for example. If you ask it to calculate a large number, it will use Python to do that.
If I ask Perplexity to simulate investment for 100 years, 4% return, putting aside $50 each month, it will use Python to write code, calculate that and then when I ask it to give me a chart it will also use python to create the image.
From GP: "But you don't use AI to define rules on the fly."
Neither Claude nor Perplexity change the rules they work by on the request to request basis. Code that Claude outputs isn't the code the Claude runs on and Perplexity did not on its own decide to create python scripts because other ways it was calculating large sums did not work well. Those tools work within the given rule set, they do not independently change those rules if the request warrants it.
Is whatever what happens between e.g. HTTP Request and Input and Output not runtime then?
1. HTTP Input
2. While (true) runAgent() <- is that not runtime?
3. HTTP Output
Additionally Claude could be triggering itself with custom prompts etc to use instances of it concurrently in parallel.
Or are you saying that the only rule is that Agent is being ran in a loop?
But the whole discussion is about how AI Agent is different from a Workflow?
The point is that workflow is that LLM is triggered in a loop ?
1) Human agent, manual retrieval (included for completion
2) one-off script to get exactly the content I want
3) Traditional workflow, write & maintain
4) one off prompt to the agent to write the script in #1, sort and arrange content for grouping base on descriptions it receives (this is what I used, 3 hours later I had a years worth of journal abstracts of various subjects downloaded, sorted, indexed and summarized in a chromadb. I’d just asked for the content, but it’s python code it left for me included a parameterized CLI with assorted variables and some thoughtful presets for semantic search options.)
5) one off prompt to the agent to write the workflow in #3, run at-will or by agent
6) prompt an agent to write some prompts, one of which will be a prompt for this task, the others whatever they want: “write a series of prompts that will be given to agents for task X. Break task x down to these components…”