String is an open-source runtime that rethinks how LLM agents interact with tools and applications. Instead of stuffing descriptions of every tool into the agent's context, String moves that knowledge into a common layer and exposes it as Markdown views on demand. Agents use just two commands: /open to see a view and /act to perform an action. On an 87-task benchmark, String achieved comparable success rates to existing methods while using 33.5% fewer tokens on completed episodes. Proper staging of information reduced wrong actions from 28% to 2%. The agent's interface stays at a constant 53 tokens regardless of how many apps are available. This matters because context window limits are a major bottleneck for LLM agents. String provides a uniform surface for apps, files, web, and shells without per-site integration, making it practical to deploy many capabilities without blowing the token budget.
Markdown interfaces cut agent context costs without reducing task success
String exposes tools and applications through staged Markdown views, loading only the actions and information an agent currently needs.
Top University
Jookyung Song · Nojun Kwak · Simyung Chang
Seoul National University · H1R.AI
Research Digest··2 min read
In one line
String provides an operating system for LLM agents where every app is a Markdown file, reducing token consumption and errors.
Before you read this
Three ideas this paper builds onagent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.150 papers →context compactionCompressing or summarizing an LLM's past context into a smaller representation to extend memory and improve reasoning within limited token budgets.71 papers →ui groundingMapping natural language commands or descriptions to specific visual elements of a user interface.23 papers →
Read it as
Song, Kwak and Chang present String, an open-source runtime that describes application views, typed actions, navigation and credentials in a single String-Flavored Markdown document.
The authors built String around two agent-facing operations: /open renders the current view, while /act validates and executes an action.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.