STRACE: Optimizing Agent Failure Trajectories
Ying Chang · hf · 2026-07-16
Proposes STRACE (Structural Trajectory Analysis and Causal Extraction) to construct higher signal-to-noise ratio contexts for long-horizon agent optimization.
- Problem: Real execution trajectories are often redundant and heterogeneous. Directly using them for reflection/optimization is inefficient and prone to overfitting, while simple truncation or sliding windows discards critical causal evidence.
- Method:
- Mines failure patterns at the batch level, filtering redundant trajectories and retaining representative failure cases;
- At the single-trajectory level, performs causal localization based on text dependency graphs, deleting non-causal steps to pinpoint the true root-cause modules.
- Results: Experiments show STRACE significantly outperforms common context filtering baselines. On the formal verification task VeruSAGE-Bench, it successfully optimized human expert-designed agents, boosting the success rate from 42.5% to 58.5%.
- Open Source: Code is publicly available.
More from coding & agent
- Claude Code often abandons the plan when a code change gets too large — roske_e · 2026-07-21
- Belgie: Embeds a TypeScript Sandbox for Python AI Agents Without Node.js — TheRealMrMatt · 2026-07-21
- Anthropic's AI Submits 65% of Engineering PRs, Slashes Prompts by 80% — Simon Willison · 2026-07-21
- OxDeAI adds signed, fail-closed authorization before AI agents can act — docybo · 2026-07-21
- Multiagent v2 playbook calls for 64 agents, diverse proof routes and adversarial checks — danshipper · 2026-07-21
- Hermes Agent adds built-in Word, Excel, PDF and PowerPoint support — Teknium · 2026-07-21