GPT-6 Astra agent beats World's Hardest Game, all 30 levels with one death

Blogger imjustnewatai published a series of posts showing an agent built with GPT-6 (Astra) completing the classic Flash game World's Hardest Game: all 30 levels cleared with only 1 death throughout, full completion in 20:51 (edited), in-game timer 10:39.075, and 30:06 of real time from level 1 to victory. The implementation and verification code are fully open-sourced.

Confirmed

• In m4, the author clarified the approach: Astra writes and supervises a harness, while the program simulates hazards, searches routes, and handles most movement automatically; Astra inspects screenshots and repairs the planner after deaths—i.e., tool-assisted control, not vision-based reflexive real-time play
• The single death occurred on level 4 (m2): after collecting a coin, the next-step search found no escape route; the fix was to require verification of a viable 120-tick (3-second) survival path before accepting a coin as a goal
• Survival-check results (m3): on the same level-4 state, simply adding the survival check yielded two simulated routes of only 60 ticks each; the corrected endpoint was just 18 pixels to the right, with the next coin reachable within 4 ticks; the author stresses this is an illustrative case, not a success-rate benchmark
• Run verification (m5): all 30 levels passed a sequential audit, with 41 game assets matching their initial hashes; the full edit retains the death footage, only trimming some long idle segments
• Open-sourced content (m6): the GitHub repo worlds-hardest-game-recorded-agent includes the harness, controller, planner, anti-cheat guarantees, methodology docs, and video verification; recording starts before any input, with no microphone or desktop capture

Why it matters

• This is a reproducible, auditable demonstration of "an LLM writing and supervising tools to accomplish high-precision tasks": the clear doesn't rely on the model's reaction speed but on Astra's ability to design the harness and repair the planner after failures
• The anti-fraud measures—hash-verifying assets, sequential audits, and record-before-input—offer a reusable verification template for similar AI game-completion demos

Timeline

• 09-18: The author published 6 consecutive threads disclosing the clear, failure-repair details, verification data, and the fully open-sourced code

2026-09-18 ~ 2026-09-18 · 6 related posts

• [source] GPT-6-built agent beats World's Hardest Game: all 30 levels, just one death — imjustnewatai · 2026-09-18
• Tool-assisted control, not reflex play: Astra wrote the harness and repaired the planner — imjustnewatai · 2026-09-18
• The one death was Level 4: agent fixed its planner with a 3-second survival check — imjustnewatai · 2026-09-18
• One change at Level 4: survival check moved endpoint 18 pixels, next coin reachable in 4 ticks — imjustnewatai · 2026-09-18
• [source] Run details: 10:39.075 game clock, 30:06 wall time, all 41 assets hash-verified — imjustnewatai · 2026-09-18
• [source] Full harness open-sourced: controller, planner, safeguards, and media verification on GitHub — imjustnewatai · 2026-09-18