DeepSeek Plugin: Verify Agent Behavior Causality with Ablation Studies
ArdiG24 · reddit · 2026-08-26
The author released a DeepSeek Harness plugin, dsh-trajectory-ablation, to address the difficulty of attributing causes to agent behavior. The plugin provides three features: Reconstruct (recreates the full context before a decision), Diff (shows changes between steps), and Ablate (conducts ablation studies). The Ablate feature removes specific context pieces and re-runs the model call to determine if that piece was the actual cause of the decision change. The author demonstrated with a real case that the tool identified a failed test output, not the original task instruction, as the trigger for the agent's code fix. The plugin is open-source, free, and requires no extra API key.
More from coding & agent
- Solo founder shares dual-prompt workflow to generate a 'Founders Guide' for projects — KennethSweet · 2026-08-26
- Perplexity launches Portable Computer local agent stack for NVIDIA DGX — AravSrinivas · 2026-08-26
- Andrew Ng maps AI engineering skills: LLM foundations, retrieval, agents, and eval-driven dev — DeepLearningAI · 2026-08-26
- Grok Bot vs Hermes: Easy assistant vs hardcore dev harness — EXM7777 · 2026-08-26
- No off-the-shelf fit: Every and Headway each build company-wide shared agents — every · 2026-08-26
- Context Sharing Remains a Bottleneck for Multi-Agent Coding Collaboration — Annual_Victory_7122 · 2026-08-26