Release Notes
Known issues and limitations
See the known issues and limitations, and learn how to circumvent them until they get fixed
Get a quick overview of the known issues and limitations for IBM watsonx Orchestrate ADK.
Issue | Description |
---|---|
orchestrate --version command | orchestrate --version command might not work or return the wrong version. To solve it, run which orchestrate to ensure it’s using your Python virtual environment (venv); if not, remove conflicting versions until it does, and if the issue persists even with the correct venv, delete and recreate the venv and reinstall the ADK. |
Tool import after upgrade | If you upgraded your watsonx Orchestrate Developer Edition and try to import a tool, you might run into permission issues. Run docker volume rm docker_tools-runtime-data to remove the existing tools and trigger a clean reinstall. |
MCP tool upload | When you upload MCP tools from the filesystem, you must zip the tool folder. You should be able to point to the root folder instead, but that isn’t working as expected. |
Renaming tools | When a tool name is modified within watsonx Orchestrate, the change is not propagated back to MCP. Currently, tool updates are unidirectional, flowing only from MCP to watsonx Orchestrate. As a result, any renaming or updates made in watsonx Orchestrate will not be reflected in MCP, potentially leading to inconsistencies or communication issues between the two systems. |
Tool flow | If you delete an OpenAPI tool from the CLI and then import the same tool again, tool flow fails. The re-imported tool gets a different tool ID, so it no longer matches the original reference. |
Tools running prematurely | Depending on the LLM you use, tools might start running before you provide all required inputs. To solve this, try a different LLM in your agent. |
First interaction with an agent might fail to call a tool | Sometimes, when you start the Developer Edition, the first message that the user sends to the agent might fail. This issue can happen because the system did not have enough time to properly start its services, or because of a pod failure. To solve this, you can try sending the same message again after a couple of minutes, after the system has enough time to restart the pod and reload the tools. |
Agent returns an empty message | In certain scenarios, asking multiple queries to the agent or even otherwise on shorter conversations, the LLM might respond with a message “None” instead of making tool calls. In such cases, try switching the agent’s style to default . |
Agent makes multiple calls for the same tool after every message | The agent might run the same tool multiple times after every interaction. To solve this, you can try providing clear instructions in your agent definition to run the tool only once, or you can try switching the agent’s model. |
Agent ignores valid tools | Occasionally, agents might return incomplete responses with invalid tool calls. This can happen if the agent recognizes the tool’s response as invalid. To solve this issue, you can try defining the tool as a YAML file, and provide a clear and more concise tool description. You can also provide clear instructions in your agent definition to interpret the tool’s responses directly and to consider the tool’s expected return data types. |
Incosistent transfer between multi-agent collaborators | When using multi-agent collaboration, agent-to-agent communication might fail due to improper instruction formatting or invalid responses. To solve this issue, you must provide clear instructions in your agent definition, explicitly detailing that one agent must transfer to another agent after completing its task. |