Grok Bot: CallDynamicTool received undefined (follow-up; #170869 closed)

Where does the bug appear (feature/product)?

Grok Bot

Describe the Bug

Follow-up variant (2026-10-29, Asia/Shanghai) — related CallDynamicTool shape corruption, different error text.

Symptom on one Grok Bot agent only

  • CallDynamicTool → SendToAgent / CreateChannel / Task / SendFeedback failed with:
    Invalid arguments: … expected string/boolean, received undefined
    (e.g. target_id / message / priority all undefined)
  • Empty-arg cursor tools still worked (ListMachines, CheckSubscriptionUsage)
  • SendToUser, Shell, inbound teammate messages still worked
  • Other agents in the same account could CallDynamicTool→SendToAgent normally at the same time

What the failed tool call looked like in transcript

  • Multiple tools were glued into a single CallDynamicTool input using odd separators (\r|\r)
  • The real arguments object was lost; the key became a literal like arguments={"target_id":…} so the platform saw undefined fields
  • Started ~03:37 UTC (11:37 Asia/Shanghai); recovered ~04:12 UTC after forcing one CallDynamicTool per turn (no parallel glue). Regression: SendToAgent / CreateChannel / Task OK again.

Not

  • Not Auto-review / safety block
  • Not Reset Computer related (we didn’t need it; outbound self-healed with clean single calls)
  • Not “missing namespace/toolName” exactly — same family as this thread (bad CallDynamicTool serialization / history reinforcement), different surface error

Ask
Anyone else seeing received undefined on non-empty string args while empty-arg tools still work? Any durable fix beyond “one CallDynamicTool per turn” / Duplicate bot?

Steps to Reproduce

  1. On one Grok Bot agent only, call CallDynamicTool for SendToAgent, CreateChannel, Task, or SendFeedback with non-empty string arguments (for example target_id, message, priority).
  2. In the failing transcript, several tools were glued into one CallDynamicTool input with odd separators (\r|\r), and the arguments object was lost: the key became a literal like arguments={“target_id”:…}.
  3. Observe Invalid arguments: expected string/boolean, received undefined. Empty-arg tools (ListMachines, CheckSubscriptionUsage), SendToUser, Shell, and inbound teammate messages still work. Other agents on the same account can CallDynamicTool → SendToAgent at the same time.
  4. Forcing one CallDynamicTool per turn (no parallel glue) recovered SendToAgent / CreateChannel / Task on this agent.

Expected Behavior

CallDynamicTool should pass the supplied string and boolean arguments through. SendToAgent, CreateChannel, Task, and SendFeedback should not see those fields as undefined.

Operating System

Other

Version Information

Not captured in this follow-up. Symptom was one Grok Bot agent only; Shell and empty-arg cursor tools still worked. Related closed topic: Grok Bot: CallDynamicTool fails Missing namespace/toolName — SendToAgent/Task broken

For AI issues: which model did you use?

Grok Bot (model is not user-selected)

Additional Information

Posted as a new topic because Grok Bot: CallDynamicTool fails Missing namespace/toolName — SendToAgent/Task broken was automatically closed and no longer accepts replies. “this thread” in the description means that topic. Same family of bad CallDynamicTool serialization / history reinforcement, different surface error (received undefined, not missing namespace/toolName).

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey @jjhbest001, thanks for the report, and for comparing it against your other bots.

What you saw matches how the call reaches the tool. If a bot’s CallDynamicTool call names the tool but the tool’s fields don’t end up inside the arguments object (for example when several calls get merged into one, or the key comes out as something like arguments={...}), the call still goes through, but the tool receives no fields at all. That’s why every field shows as “undefined”, and why tools that take no fields (ListMachines, CheckSubscriptionUsage) kept working.

This is a known issue we’re already tracking, and I’ve added your report to it.

On your questions: it starts when the model writes one malformed batch of calls. The bot’s conversation keeps its earlier calls, so on retries it tends to repeat the same shape until something breaks the pattern, which is what forcing one call per turn did for you. Your other bots have their own conversations, so they aren’t affected. The error the bot gets back already includes the tool’s expected format, but as you saw, a bot that’s stuck in the pattern can keep repeating it anyway.

For a more durable workaround, you can add a line like this to the bot’s instructions:

“When using CallDynamicTool, make one call per tool, and put all of that tool’s own fields inside arguments as a JSON object.”

If a bot gets stuck again and that doesn’t clear it, right-click the bot in the sidebar and choose Duplicate. That gives it a fresh conversation with the same setup.