Rearchitecting Letta's Agent Loop Recent

Published: October 14, 2025 | Original Post

---

Lessons from ReAct, MemGPT, and Claude Code in designing the agent loop.

What is Agent Architecture?

At each iteration, agents must decide:

• Call tools or send messages
• Reason (allocate compute for thinking)
• Terminate the loop

These decisions can be controlled by the LLM itself or by the agent architecture managing invocations.

Evolution

ReAct (Early)

Sequential reasoning and actions through formatted prompts -- predating native tool calling and reasoning capabilities.

MemGPT

The first stateful agent with persistent memory. Built on tool calling, it used special keyword arguments (thinking, request_heartbeat) to manage reasoning and control flow.

Letta V1

The new architecture:

• Leverages native reasoning from modern models
• Removes tool-calling requirements for reasoning
• Simplifies system prompts

Trade-offs:

• Loses prompted reasoning for non-reasoning models
• Eliminates heartbeat functionality
• Reduces tool rule flexibility

The Reasoning Token Dilemma

Modern frontier models encrypt reasoning tokens, preventing developer access. Prompted reasoning offers transparency but performance comparisons remain unclear across use cases.

Recommendation

Use the new Letta V1 architecture for latest reasoning models like GPT-5 and Claude 4.5 Sonnet.