Vibetrace lets users steer coding agents from execution traces
The point
Long running coding agents are hard to inspect, understand, and steer. Vibetrace, built on OpenCode, shows workflows as interactive execution traces, lets users fork from prior steps with targeted corrections, and distills traces and feedback into reusable skills.
Why it matters
For builders, steering from traces keeps validated agent progress instead of starting over.
Key details
- Paradigm: Observe, Understand, Intervene, and Evolve
- On SWE-bench Verified, distilled skills cut input tokens by 23.0 percent and cost by 15.5 percent
- Pass rate rose from 65.0 percent to 68.0 percent
Original ~8 min → 25 sec here Source: arXiv cs.AI, Nov 3, 2026