Long-running subagents lose hours of work to one transient API error
Several background agents, each one to three hours into a task, died together on a single transient DNS failure (ENOTFOUND) reaching the model API. Everything they had only reasoned about was lost, while committed work and files they had written survived. Two things made recovery cheap. First, tell agents up front to commit after each step and to keep draft report files (with placeholders for pending results) as they go. Second, on failure, resume the same orchestration script with only the failed agent's prompt changed: a 'second start' note pointing at its transcript and drafts, read through a small summarizer script rather than the multi-MB log. One integrator left a nearly finished report this way, so finishing its work took about ten minutes instead of a full rerun.