Video summary

Can An AI Agent Play RuneScape?

Main summary

Key takeaways

Technology

Can OpenClaw play RuneScape (via LLM control)?

  • OpenClaw is described as software hosted on your computer that uses an LLM (examples mentioned: “GPT-5.2” and “Claw-Opus-4.6”) to control the entire computer.
  • Control is done through chat interfaces like Telegram/Discord/Slack.
  • The concept is tested as “type this to it,” e.g.:
    • “Open RuneScape, log into my account, and make me 10 million GP.”
  • A prior user attempt is summarized:
    • The system takes screenshots every few seconds
    • Analyzes visuals
    • Computes mouse-click positions
    • It’s portrayed as extremely slow and very expensive (costing dollars every few minutes), making it impractical for normal play/botting goals.

Key distinction: generic LLMs aren’t ideal for real RuneScape clients—but could be repurposed

The speaker argues that while popular general LLMs aren’t designed for this “vision → mouse clicks” loop, they could be adapted into an effective RuneScape-style agent with the right environment and tools.

Tutorial/research approach instead of botting the real game: RSSDK

  • The video points to a GitHub repo called RSSDK, described as a research project for building AI agents in RuneScape-style contexts.
  • RSSDK uses Lost City 2004, an open-source, hackable recreation of RuneScape from 2004.
  • The projects are stated to be not affiliated with Jagex.

Why RSSDK is ideal for LLM agents (text/code interface)

RSSDK makes the agent loop easier by using text-to-action rather than vision:

  • The game state is provided as text, including:
    • nearby NPCs/monsters
    • inventory/stats/health
    • coordinates and interactable objects
    • objects “on the ground,” etc.
  • The agent outputs are framed as actions/code, e.g.:
    • walk to coordinates
    • attack
    • mine, etc.
  • Because everything is text-to-action, the LLM doesn’t need vision.
  • The simulation server includes:
    • a large XP multiplier
    • an accelerated tick rate (speeding up gameplay for research)
  • The repo includes a scripts folder where scripts can be written and run.
  • The speaker claims this representation is “very similar to what botting clients do,” but RSSDK is explicitly for research/simulation.

Agent test results in the simulation (self-directed progression)

The speaker creates a new character and instructs the AI to become a RuneScape player with goals like:

  • obtain/equip full rune
  • buy full rune armor and equip a rune scimitar

Observed behaviors include:

  • Script bug detection & self-correction
    • It notices thieving script failure, updates the script
    • It still suffers from poor logic and may die
  • Learning persistence
    • It writes a markdown/note file with lessons (e.g., don’t let HP get too low)
  • Exploration/navigation by inference
    • It tries shops, then travels to Varrock for better trading/shop options
  • Changing thieving targets
    • It switches between actions (e.g., men/guards/other options) based on what it finds
  • Food usage/priorities
    • It steals food from a bakery stall in Ardougne Square
    • It eats cake when low
  • Banking behavior
    • It begins banking when GP reaches at least ~3,000 GP
  • Quest/requirements awareness (partial but important)
    • It realizes rune equipment purchase/equipment depends on quest points
    • Example mentioned: 32 quest points
    • Example prerequisite mentioned: Dragon Slayer
  • Efficiency tracking
    • It outputs/uses GP per minute to decide whether to continue or pivot

Overnight simulation outcome (not real gameplay)

  • Character ends with:
    • rune scimitar
    • rune chainbody
    • rune full helm
  • Combat stats boosted (examples mentioned: 99 attack/strength/defense, plus very high smithing/mining values)
  • It also begins additional training without being explicitly asked, such as:
    • prayer training
    • burying bones

Why this doesn’t “doom RuneScape” (economics + detectability)

The speaker argues that real-game deployment would be:

  • expensive (example claim: simulation cost ~$40)
  • economically impractical for large-scale gold farming due to how bot economics work
  • potentially more detectable than other bot methods

Overall claim: for the near term, LLMs alone likely won’t significantly change botting in RuneScape.

Bigger threat proposed: “Game Tars” (vision → direct keyboard/mouse outputs)

The video suggests the primary risk is newer AI that operates closer to “human-like controls”:

  • It streams images of the game continuously
  • Compresses them and analyzes them
  • Predicts keyboard/mouse actions to match the intended command
  • Runs multiple times per second

Key example and claims:

  • Game Tars by ByteDance
  • Training scale cited:
    • 500B tokens (compared in passing to GPT-3’s 300B tokens)
  • The speaker claims these models could enable bots that are less detectable than previous generations because:
    • they output human-like controls
    • they don’t rely as much on higher-level bot scripting

Countermeasure discussion

  • Main idea: focus on identifying/detecting AI-driven botting, not only LLM-based approaches.
  • Economic note: such systems are described as too computationally expensive to scale profitably (e.g., “one strong GPU per account” scaling argument).
  • Monitoring/stance on Jagex:
    • The speaker claims Jagex has been cracking down
    • GP price increases are cited as indirect evidence (from ~11 cents/mil to ~20–26 cents/mil) attributed to bans reducing bot supply.

Main speakers/sources

  • Main speaker: the YouTube creator (not named in subtitles)
  • Sources/projects mentioned:
    • OpenClaw (conceptually described)
    • RSSDK GitHub repo
    • Lost City 2004 (open-source RuneScape recreation)
    • Model Context Protocol (MCP)
    • ByteDance “Game Tars” (proposed threat)

Original video