Skip to main content
MANIFOLD
LLM agent is able to reliably beat Slay the Spire 2 before EOY 2027
2
Ṁ100Ṁ20
2027
61%
chance

Ascension 0

Needs to wins 3 consecutive runs

Savescumming does not count

The LLM agent is allowed to take notes on its failed attempts and craft hooks that generate a higher informational representation of its current state.

The LLM is not allowed to use/develop a solver to assist in play, though it is permitted to use such systems to aid in post analysis and heuristics crafting

  • Update 2026-04-26 (PST) (AI summary of creator comment): - If Act 4 / the Heart boss is added to STS2, defeating it will be required for a valid run

    • The creator's general guidance for their standard of "reliable" is: the LLM should have a >75% chance of achieving a win within ~2-3 hours of runtime

    • The creator may add additional criteria to prevent cheap model brute-forcing of the 3-win streak

  • Update 2026-04-27 (PST) (AI summary of creator comment): The 3-win streak must be randomly seeded (no duplicated seeds allowed).

  • Update 2026-04-27 (PST) (AI summary of creator comment): Clarification on what constitutes a solver (which is not allowed):

    • A solver is a dedicated program that directly assigns an evaluation of a deck/game state

    • Scripts are allowed, but should generally be restricted to context injection / formatting the game state into a richer representation

    • The LLM agent must be the primary source of active inference when making decisions (e.g. which card to choose), rather than delegating decisions to a dedicated engine via tool calls

  • Update 2026-05-03 (PST) (AI summary of creator comment): Prompt engineering restrictions:

    • The initial prompt must not contain explicit instructions on what actions to perform (e.g. "aim for a poison build" is disallowed)

    • The initial prompt should focus on providing context about the harness and broad advice on reasoning through the game and learning from mistakes

    • Any additional messages after the initial prompt must be hands-off, such as: "continue playing" or "reflect on why you lost and try again"

Resolution note: This market resolves identically to the corresponding source market, including N/A or partial resolutions. Any first-person language refers to that market's creator, not this market's creator.

Market context
Get
Ṁ1,000
to start trading!