I noticed the SDK supports a `system_message_suffi...
# proj-agent-canvas
a
I noticed the SDK supports a
system_message_suffix
on Agent Profiles, but it doesn't seem to be exposed in the Agent Canvas UI. It seems like a useful addition since it would let users create specialized profiles (coding, research, security review, documentation, etc.) without changing the underlying agent. Was this intentionally left out, or would it make sense to expose it?
s
The SDK currently supports more features than what the canvas uses • system_message_suffix • skills per agent profile • mcps per agent profile • condenser per agent profile • ... The canvas has not caught up yet. It's "trivial" to connect the canvas to the sdk backend, but it's really a UX/UI decision
About
system_message_suffix
I do have some doubts about the utility of it. LLMs are trained to be agent engines, so I guess they are "used to" having a general system prompt and they should be great at specializing given the first specific user message. So it's a nice to have but I wouldn't expect a big diff of
system_message_suffix
vs
suffix + first user message
... I might remove it from the SDK for simplicity
a
I would like to tune the prompt and have things like skills/mcp per agent profile. It's also convenient for me if the agent already knows I want to just do research and I don't have to tell it at teh start of every interaction. (So I like it having the suffix there -- or have another way to get similar behavior)
w
if we were to expose
suffix
, it'd probably persist across ALL sessions though...otherwise you might as well put it in the prompt
but it would be nice to customize the agent's behavior with things like "Be as concise as possible, always use bullet lists when sending messages"
a
exactly, my goal is just to customize agent's behavior and a good starting point is a nice box in the UI that I could add additional instructions like you suggested it. How exactly we implement I am flexible. This would add a bit more substance to the idea of an agent profile.
nod 2
☝️ 1
h
I think that is the exact story some users want with "Agent Profiles"
s
well, the SDK supports it! so it's just about figuring out what UX we want in canvas for this 😅 we can fit a small section for per agent suffix somewhere
w
nice yeah, I'd love to add a text field to each agent profile 🙂