---
type: "post"
id: "knowing-how-is-not-knowing-what-14-minute-edc126"
slug: "knowing-how-is-not-knowing-what-14-minute-edc126"
title: "Knowing How Is Not Knowing What (14 minute read)"
description: "Yunze Xiao published an essay on RL environment design for computer-use agents."
retrieval_nugget: "Yunze Xiao published an essay on RL environment design for computer-use agents."
published_at: "2026-10-26"
updated_at: "2026-10-26"
record_date: "2026-10-26"
date_kind: "published_at"
topics: ["agent-loops","automation"]
entities: ["Yunze Xiao"]
related_patterns: ["goal-scoped-loops-replace-manual-continuation","harness-architecture-outlives-model-choice"]
novelty_label: "new_angle"
what_changed: "It argues RL environments should test organizational know-how and goal achievement, not just software operation—an extension of prior critiques of benchmark design and agent evaluation."
prior_refs: ["memory:431f2dd3-0c91-81a0-afd7-0b71696c42c6","memory:495e0c62-b772-84b6-a217-ab56c94c540b"]
source_url: "https://algoroxyolo.github.io/blog/2026/knowing-how-is-not-knowing-what"
source_urls: ["https://algoroxyolo.github.io/blog/2026/knowing-how-is-not-knowing-what"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:knowing-how-is-not-knowing-what-14-minute-edc126"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/knowing-how-is-not-knowing-what-14-minute-edc126/","markdown":"https://newruntime.com/posts/knowing-how-is-not-knowing-what-14-minute-edc126.md","json":"https://newruntime.com/posts/knowing-how-is-not-knowing-what-14-minute-edc126.json"}
---

# Knowing How Is Not Knowing What (14 minute read)

## Retrieval answer

Yunze Xiao published an essay on RL environment design for computer-use agents.

Yunze Xiao published an essay on RL environment design for computer-use agents. It argues that executing a software procedure should not be treated as equivalent to completing an organizational task. In the examples, sending a headcount request does not secure an authorized position, and submitting an $8,000 refund does not satisfy a policy requiring a second approval above $5,000.

The proposed environment includes more than application interfaces and task instructions. Xiao calls for persistent organizational state: roles, authority, budgets, deadlines, documents with cross-references, messages, and domain rules that may be unwritten. Actions should change that state across tool calls. Evaluation should inspect the resulting world, including whether an authorized approval occurred and whether required resources exist.

The essay also proposes treating real company records as raw material rather than directly reusable environments, because archives may contain personal data, cannot readily be reset or varied, and record past events rather than a counterfactual definition of correct work. It does not release a benchmark, dataset, environment, model, or experimental results establishing that this approach improves agent performance.

Source: https://algoroxyolo.github.io/blog/2026/knowing-how-is-not-knowing-what

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Loops](https://newruntime.com/topics/agent-loops.md): Explore the agent-loops topic hub.
- topic hub: [Automation](https://newruntime.com/topics/automation.md): Explore the automation topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)