Codex + Astra xhigh tackle MazeBench: static BFS solver nets 44 gems in ~2 hours
xeophon · x · 2026-09-12
FakePsyho shares progress on MazeBench using codex with astra xhigh:
- Setup: the agent gets full source-code access (no visual component), making it substantially easier than the official test; prompting was minimal — the author only outlined the approach, then went hands-off.
- Approach: a fixed, optimized BFS solver for individual rooms plus a global solver that stitches solved rooms together.
- Results: a fully static solver earned 36 gems on 5.6-sol in roughly 2 hours on a single CPU; Astra's static solver reached 44 gems in similar time, partly aided by resuming a run that had stalled at 44 gems.
- Takeaway: an exploration of how far a minimal-prompt coding agent can go by orchestrating classic algorithmic solvers.
More from coding & agent
- Running multiple persistent agents on a single VM works surprisingly well — remilouf · 2026-09-12
- Codex adds manual confirmation pauses, sparking criticism over long-task usability — doodlestein · 2026-09-12
- OpenBMB open-sources MiniCPM5-2B, a 2B model that tops agentic benchmarks under 4B — SimplyAnnisa · 2026-09-12
- Dev hits another suspect Codex rate-limit pause, delegates sessions to a second agent — doodlestein · 2026-09-12
- Dev's multi-agent setup: Telegram topic groups plus ephemeral agents from GitHub issues — granawkins · 2026-09-12
- How one dev keeps 3 AI agents busy in parallel with email-length prompts — granawkins · 2026-09-12