Is OpenAPI enough for tool use?

[yagnyaPatel]

I really haven't found it good enough just due to the high information density with many irrelevant specifics. Even harder than the context size limit is the context stuffing problem, where information-dense context leads to poor LLM accuracy. Some of the workarounds to the tool orchestration issue we've seen are:

• Hub and Node toll calling delegation
• OpenAI Swarm, CrewAI
• Hardcoded workflows with agentic triggers

On this basis, we've been building agents.json to enable a more interoperable and deterministic way for LLMs to interact with APIs. Would love to get the AX community's thoughts and work on this with AX community.

What are the current approaches/workarounds you've been able to rely on for tool calling?

[abhiaiyer91]

100% agree.

Without more work an LLM does not just use your tool.

What has worked:

• Really clear tool names and descriptions. The description should be rich and as close as to the description limit you can.
• Then adding more tool descriptions to the system prompt to reinforce how a tool works, when to use it.

Both of those put together and I can command an LLM to do things mostly well!

Have you seen the something like this?

[sean-roberts]

At a certain point of context (# of tools usually) the "Hub and Node toll calling delegation" seems to be the best approach I reach for. (I'd also be curious about your thoughts on this discussion)

With agents.json, what is it that solves this problem? It would seem to me the quality of selection is directly correlated to the quality of description, is that still a factor with agents.json but you also get a predictable format (which I know is for the interoperability aspect).