Observed arrival · 2026-10-11
GATC-S: A step-by-step dataset of coding-agent sessions
GOAT labs presents 4,317 recorded sessions from frontier-model coding agents, preserving their prompts, replies, tool calls, and failures.
- For
- Researchers studying frontier-model coding agents
- Worth noticing
- The 100-session public sample is balanced across four model families; the page reports 5,026 failed calls among 136,482.
Field notes
Each record preserves ordered user prompts, agent replies, tool calls and arguments, recorded results, and failed calls alongside subsequent steps; reasoning and token counts appear where logged. The page reports a median session length of 20 steps, with one in ten reaching at least 128, and says 2,296 sessions record touched files. Its dry-run coding example is labeled an illustration, not a dataset session. The public sample contains 100 sessions balanced across four model families.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue