https://www.boundaryml.com logo
I'm thinking, if future Agents primarily
# general
v
I'm thinking, if future Agents primarily use CLI tools as the main integration method for tools, then tool calls only need to integrate a very limited number (such as simple ones like ReadFile/EditFile). In that case, whether BAML natively supports function calls isn't that important anymore; the main output just needs to be Bash
Of course, I think if the trend continues, BAML's advantage will also diminish, because iterations can be made through Bash feedback without needing BAML's SAP to ensure tool calls. Current models are also becoming increasingly capable in handling Bash commands.
g
It’s an important perspective to stay aware of! I personally still believe there is a lot of power in specialized tools. Bash is indeed the most important tool for computer-use agents (including coding agents). But for most users-facing agents in apps, it’s valuable to give your agent tailored tools for updating app state, and the competitive advantage will go to the app with the best combination of state management, context management and LLM testing. I don’t see the Bash-is-all-you-need trend continuing indefinitely, IOW. What do you think? 🙂
w
If you mean this post:https://www.reddit.com/r/LocalLLaMA/comments/1rrisqn/i_was_backend_lead_at_manus_after_building_agents/?share_id=EClJw68jaqptf0ibMxMQW&utm_name=iossmf&seeker-session=true The original post has mentioned: > CLI isn't a silver bullet. Typed APIs may be the better choice in these scenarios So I think BAML can go well too.
v
Thank you for sharing. This is my first time seeing this post, but these scenarios are not the main ones I work with.
Copy code
Boundaries and limitations
CLI isn't a silver bullet. Typed APIs may be the better choice in these scenarios:

Strongly-typed interactions: Database queries, GraphQL APIs, and other cases requiring structured input/output. Schema validation is more reliable than string parsing.

High-security requirements: CLI's string concatenation carries inherent injection risks. In untrusted-input scenarios, typed parameters are safer. agent-clip mitigates this through sandbox isolation.

Native multimodal: Pure audio/video processing and other binary-stream scenarios where CLI's text pipe is a bottleneck.
Can BAML's new language features handle the third scenario?
p
I read that same reddit thread and was wondering if skills can be converted to baml. Write a CLI wrapper over these and allow the agent/harness to pick the baml command(s) to run?
g
@vEth Can Boundary's new langugae handle native multimodal workflows? Yes-depending. The one caveat is that we don't have a good story for multimodal LLM outputs yet. However I'd answer "yes" - any general purpose language (including both BAML and bash!) has facilities for keeping the stdin/stdout path low-traffic and dealing with images in a side-channel.
I think this video explains why Bash isn't everything. I also wonder whether BAML can solve the pain points mentioned here. @User