---
title: Agent framework research
description: Five shares — a repo, a ChatGPT thread, a talk, a paper, an X thread — compiled into one working view of the agent-framework landscape.
source: ManticsCore
kind: clip
url: https://manticscore.com/clip/agent-framework-research
---

# Agent framework research

Five shares — a repo, a ChatGPT thread, a talk, a paper, an X thread — compiled into one working view of the agent-framework landscape.

# Agent framework research
5 sources · one thread · compiled from a ManticsCore session

The question behind this session: which agent architecture actually fits an iOS-first product — a typed tool-call core, or a looser prompt-orchestration layer? Everything below was shared in from the iOS Share Sheet over two days; ManticsCore connected the sources and kept the thread.

- GitHub repo: A TypeScript toolkit for tool-using agents — typed tool-call core, streaming runs.
- ChatGPT conversation: Working through failure modes of free-form tool prompting.
- YouTube talk: A conference session on agent evals and why loops beat chains.
- PDF paper: Tool-use benchmarks across schema-typed vs free-text calling.
- X thread: Practitioner notes on shipping agents to production.

Where it landed: typed tool-calls win for anything user-facing — the schema is the contract, and the contract is what makes streaming UI possible. The looser orchestration patterns are better treated as internal plumbing than product surface.

Full interactive version: https://manticscore.com/clip/agent-framework-research