A self-devised 'freeze a service first' method let Codex pass a Terminal-Bench task with zero public solves
Present-Quantity-813 · reddit · 2026-08-19
The author wrote a method and handed it to Codex, which passed the ico-path-patch task on Terminal-Bench 3.0 — a task with 59 public runs across 11 model/agent configurations and zero passes. The same model used (gpt-5.6-sol at max reasoning effort) went 0-for-5 in the public record. One run scored 19/19 on the official verifier within the 90-minute limit.
The method: the agent doesn't start on the task. It first builds itself a small service for the task, freezes it, then works through that service instead of re-deriving constraints every few turns.
Motivation: on long runs, an agent's picture of where it is drifts from where it actually is. The author derived this drift theory from running a multi-agent system, after one agent spontaneously used it to diagnose its own drift.
Caveats: the author passed on the 4th scored attempt, and there's no control run with a build phase but no method text — so "any build phase would do" remains a live alternative explanation. The author invites others to reproduce with their own stacks or his free write-up.
Related event: Codex First to Pass Previously Unsolved Terminal-Bench Task(2 posts)→
More from coding & agent
- Apple code signing drags for hours; official advice: just wait a week — pjausovec · 2026-10-16
- Hermes: a 4B model with 8K context got worse with more test-time compute, pointing to contextual reasoning — mangahomanga · 2026-10-16
- Flux 3 Klein's Default Templates Are Broken Without Its Prompt Upscaler — StrictMove5839 · 2026-10-16
- 21-person team ships hundreds of PRs weekly with AI writing 97% of the code — rseroter · 2026-10-16
- ClosedMCP: Set Up MCP Servers Once, Share Them Across Your Whole Team — Ok_Sprinkles1301 · 2026-10-16
- Heisentools: let an agent rewrite a mock.py until it mirrors any real tool's API — kieranklaassen · 2026-10-16