Starstruk Logo MarkStarstruk

typescript_1.1.0

2FastLabs/agent-squadtypescript_1.1.0Jul 1, 2026by cornelcroi

AI Summary

Introduces the GroundedAgent pattern to mitigate LLM hallucinations by separating tool execution (gatherer) from response generation (presenter).

Key Highlights

  • Two-LLM architecture: Gatherer calls tools, Presenter generates replies based solely on curated tool output.
  • New classes: GroundedAgent, ToolOutputCurator (DataBlockCurator, PerToolCurator), PresenterPrompt.
  • Includes runnable example: grounded-agent-chatbot.

New Features

  • GroundedAgent implementation
  • ToolOutputCurator system
  • PresenterPrompt configuration

Full Release Notes

## GroundedAgent (two-LLM anti-hallucination pattern)

Adds `GroundedAgent`: a gatherer LLM calls tools and sees the raw results but never speaks; an isolated presenter LLM writes the reply from **only** the curated tool output, so it can't invent values beyond what was fetched. A no-tool turn is answered by the gatherer directly, skipping the presenter.

- `GroundedAgent` / `GroundedAgentOptions`, `ToolOutputCurator` (`DataBlockCurator`, `PerToolCurator`), `PresenterPrompt`
- Docs: Built-in Agents → Grounded Agent
- Runnable example: `examples/grounded-agent-chatbot`

Feature merged in #551.