ReAct Orchestration

Kernel Service uses the ReAct (Reason + Act) pattern for task execution. The agent alternates between reasoning about what to do and taking actions via tools.

The Loop

1. Build context (agent + objective + tools)
2. Call LLM with available tools
3. If LLM returns tool calls:
   → Execute tools
   → Add results to context
   → Go to step 2
4. If LLM returns final answer:
   → Record output
   → Complete task

Safety Guardrails

• Max iterations: Hard limit (default 10) prevents infinite loops
• Tool permissions: Agents can only use tools they’re explicitly allowed
• Execution modes: Outbound actions (email, messages) are draft-only by default
• Budget enforcement: Token/cost limits per tenant (coming soon)

Multi-Provider Fallback

If the primary LLM fails, the orchestrator automatically falls back:

Tier	Primary	Fallback 1	Fallback 2
Cheap	Gemini Flash	Claude Haiku	GPT-4o-mini
Expensive	Claude Sonnet	Gemini Pro	GPT-4o