Add Option to Preserve ReAct Reasoning ("Thought") in Context

Problem

When using ReAct-style agents in multi-agent configurations, the system removes "Thought:" portions from responses before storing them in context. This causes format inconsistency and behavior instability.

Current Behavior

Child agents output: Thought: ... Collect: ...

System stores only: Collect: ... (question text)

Next LLM requests receive plain text instead of ReAct format

This contradicts Instructions that require ReAct format

Impact

Format Inconsistency: LLM receives conflicting signals (Instructions require ReAct format, but context shows plain text)

Behavior Instability: Agents fail to maintain ReAct format in subsequent turns

Multi-turn Degradation: Problem worsens over multiple conversation turns

Production Risk:

Non-deterministic behavior in enterprise workflows

Increased difficulty in debugging agent decisions

Reduced trust in agent outputs

Root Cause

Per Support case, this is intentional for:

Token optimization (reduce context size)

UI display (show only user-facing content)

However, this breaks ReAct agents because:

ReAct agents depend on context consistency

LLMs often prioritize recent conversational context over system instructions when inconsistencies exist

Modified context disrupts the reasoning chain

Proposed Solution

We propose adding a configuration option such as:

preserve_reasoning_in_context: true

When enabled:

Store complete agent outputs including "Thought:" in context

Maintain ReAct format consistency

Apply filtering only at UI presentation layer

Rationale:

Some users prioritize token efficiency (current behavior)

Others require reasoning consistency for correctness

This option enables both use cases without breaking existing deployments

Benefits

Stability: Consistent output format across turns

Reliability: Predictable multi-agent behavior

Flexibility: Choose between token optimization and format stability

Backward Compatible: Default behavior unchanged

Use Cases

Multi-turn information gathering

Complex multi-agent workflows

Production environments requiring stable behavior

Scenarios prioritizing format adherence over token optimization

Alternatives Considered

Instruction-only control: Insufficient (LLMs prioritize context)

react_intrinsic style: Requires complete redesign

Stronger Instructions: Cannot overcome contradictory context

Notes

Current implementation ties together display, storage, and LLM input.

This enhancement can be implemented by:

Introducing a flag in the context persistence layer

Applying conditional filtering before UI rendering (instead of before storage)

Solution should clearly decouple:

LLM context storage (full fidelity, including reasoning)

UI presentation layer (filtered, user-facing only)

while maintaining backward compatibility.

This enhancement would improve reliability of ReAct-style agents in watsonx Orchestrate, especially for complex multi-agent scenarios in production environments.