GumossPal Agent
PALWORLD. PLAYED BY AN AI.

It lost to Grizzbolt.
Then it came back.

Meet Pal Agent, an AI playing Palworld. Watch it fight, dodge, and take another shot at the tower.

Straight to the good bits
FIRST TOWERZoe & Grizzbolt
RECORDED GAMEPLAY · 20 OCT
REAL GAMEPLAYAI AT THE CONTROLSMISTAKES INCLUDED

Earlier defeat. Later victory. Different runs, teams, and gear.

SKIP TO THE GOOD BITS

Bad calls. Close calls.
One very good ending.

Pick a clip. Watch the decision play out.

Selected moments from the 19 and 20 October attempts. Gameplay at original speed.

FROM THE WINNING RUN
BACKSEAT GAMING ENCOURAGED

An attack is coming.
Your move?

You’re drawing the bow. Grizzbolt is closing in. What would you do?

Choose an option, then click the video to play. These recorded examples come from different moments; your choice doesn’t change the outcome.

From taking the hit to dodging it: what changed between attempts?

Behind the scenes

RSI · WHAT CHANGED BETWEEN ATTEMPTS

First, it took the hit. Later, it dodged.

Between attempts, AI coding agents review footage, revise instructions and controls, then test again. These two tower runs show the behavior that process was working to improve.

EARLIER ATTEMPT · 10 SEC

Still drawing. Already getting hit.

Watch the bow stay up as the electric attack lands.
LATER ATTEMPT · 12 SEC

Dodge first. Then take the shot.

Watch it dodge during the attack, then return to the bow.
THE CHANGE BETWEEN RUNS

Let danger interrupt the shot.

AI coding agents revised the instructions and controls so KEV could prioritize an available dodge before charging another shot. The loop below shows how that change was reviewed and tested.

HOW THE RSI LOOP CONNECTS THE TWO

  1. Review

    AI inspects the failed attempt and action logs.

    Evidence: the hit lands

  2. Diagnose

    Why did the danger cue fail to interrupt the bow attack?

    Problem: attack priority

  3. Revise

    AI edits instructions and controls to prioritize the offered dodge.

    Change: dodge before charging

  4. Test again

    Run the revised build. Inspect its decisions and the result.

    New run, new evidence

↶ AI reviews the revised system’s next run—and uses it to guide the next revision.

These attempts used different teams, levels, and gear, so the comparison does not isolate the effect of one change. RSI here updates code and instructions between runs; model weights stay unchanged.

People set the goals and limits. Pal Agent reads game data through a mod; it does not rely on screen images alone.

RECORDED GAMEPLAY

Pal Agent in one minute.