It pretty clearly means AI that does things, as opposed to just presenting things. There can be a debate about the technical word "agent" in swe, but i dont see the confusion in the particular context of gen AI. Same for things like MCP, why does everyone suddenly have a stroke and argue over things with clear meaning now?
The title seems like clickbait but also reflects an annoyance/frustration with overuse as marketing/sales tends to do. I ignore that and ask what exactly does your product do, then decide how to categorize it.
My hand-wavy definition is basically as a chatbot gives an AI multi-modal input and output to a user, an agent has 'hands' which can produce an effect on a system rather than reporting to the user directly via media. The method of integration isn't really important, it could be using a public network API or as part of a compiled static binary, whatever.
When “it becomes challenging to benchmark performance and ensure consistent outcomes,” should I worry that different products don’t fully comprehend client-server architecture?
Barely anyone will use this term in 2 years, just like nobody talks about copilots anymore. But for now: everything is labelled as agent.
Microsoft literally renamed their Office Suite to more prominently feature Copilot.
Facebook is now Meta and no one cares about the supposed titular product anymore.
And in two years they'll drop it!
Remember cortana!
Tay Tweets remembers.
It pretty clearly means AI that does things, as opposed to just presenting things. There can be a debate about the technical word "agent" in swe, but i dont see the confusion in the particular context of gen AI. Same for things like MCP, why does everyone suddenly have a stroke and argue over things with clear meaning now?
It's the blast processing that enables clipped to say "It look like you are writing a letter"
The title seems like clickbait but also reflects an annoyance/frustration with overuse as marketing/sales tends to do. I ignore that and ask what exactly does your product do, then decide how to categorize it.
My hand-wavy definition is basically as a chatbot gives an AI multi-modal input and output to a user, an agent has 'hands' which can produce an effect on a system rather than reporting to the user directly via media. The method of integration isn't really important, it could be using a public network API or as part of a compiled static binary, whatever.
There is of course the "Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents"[0] paper.
[0] https://arxiv.org/abs/2405.02957
Anthropic blog has a nice distinction between LLM Workflows and Agents:
LLM Workflows: predefined code paths. Includes retrieval/tools/memory, routing based chatbots, orchestrator-worker, iterative evaluator etc
Agents: dynamically direct themselves. reasoning and planning, typically using tools based on env feedback in a loop.
Have we struggled with defining what a daemon is?
When “it becomes challenging to benchmark performance and ensure consistent outcomes,” should I worry that different products don’t fully comprehend client-server architecture?
A daemon or “non deterministic data pipeline” doesn’t command grandiose valuations. It must sound like magic.
Some terms are all encompassing. This is one. Similar to the components or monad arguments.
They invested billions into AI so you can be damn sure they call everything AI.