LLM passing wrong parameter while performing tool calls

[Samgh1011]

I’m using a multi-agent setup (AgentOS → Team → Agents → MCP tools).

Each request contains a Context block with SOME mandatory params that must be passed in every tool call.

Current Approach

• Instructions force the system to:
  • Read Context
  • Copy all MANDATORY params exactly
  • Pass full input during delegation
• Additional instructions + skills define tool usage as well

Example Request

curl --location 'https://localhost:8005/api/agui' \
--header 'accept: application/json' \
--data-raw '{
  "messages": [
    {
      "role": "user",
      "content": "User prompt: <action>; Context: { Param1: ..., Param2: ..., Param3: ..., Param4: ..., Param5: ..., Param6: ... }"
    }
  ]
}'

Issues Observed

❌ Params missing WHILE MAKING TOOL CALL as team forget to delgate the context section to agent even after telling clearly that in the instrction of teams/ or agent passing hallucinated values for these sensitive fields 

❌ Wrong values passed

❌ Delegation sometimes skips context


Core Problem

LLM is responsible for copying + injecting these params → non-deterministic behavior

Goal

Ensure these 6 params are:

Always correct

Not constructed by the LLM

LLM should only decide intent/tool, and the other arguments from user query , BUT NOT THE CRTICAL  sensitive args. 

Question

Has anyone handled this scenario where SENSITIVE params are passed via context but LLM makes them unreliable?

What’s the best way to guarantee these params are preserved correctly without depending on the LLM?

IS THERE A WAY BY WHICH AGNO LETS THE PART OF THESE SENSITIVE PARAMS TO BE PASSED MANUALLY , rather than having to depend on LLM for this , for non sensitive args let it be constructed by the LLM itself.

[omni-front]

it sounds like you're dealing with an issue where the llm is unpredictably handling sensitive parameters. to fix this, consider separating the handling of sensitive parameters from the llm's responsibilities. instead of relying on the llm to pass these critical params, you can set up a mechanism where these params are manually injected into the tool calls after the llm determines the intent or tool needed. this way, you maintain control over the sensitive data and only use the llm for the less critical parts of the request.

[Samgh1011]

@omni-front , can u help me with a small starting script ?