Why unattended agents redo finished work: a thought experiment on task completion state
ClickOk5811 · reddit · 2026-10-09
A Reddit post frames a thought experiment for a common long-run agent failure mode: imagine a coworker who can never cross finished tasks off a list of 40 jobs — by job 25, they'd re-examine and likely redo completed work.
That, the author argues, is what an unattended agent does on long multi-step runs, like a refactor that renames an interface and updates every call site, then re-touches files it already got right hours later. Nothing marks steps as "closed," and every finished tool call sits in context weighted like the current step.
The problem shows up less with a human in the loop because each reply implicitly tells the model prior steps are settled; remove the human and nothing takes over that role. A bigger context window won't fix it — it's just more room for finished work to linger. The author points to explicit mechanisms: a task list the agent must update before moving on, or collapsing finished steps into one-line notes. Untested; the author asks overnight-job runners what has worked for them.
More from coding & agent
- Swapping harness lifts GPT-5.6 repo migration from 6.5% to 31%, paper finds — omarsar0 · 2026-10-09
- Developer builds wildfire analysis tool with Opus 5.5: fire progression and hot spot tracking — natesiggard · 2026-10-09
- Subagent-as-a-service: betting on 'expert' agents like Harvey as tool calls for vertical AI — nbaschez · 2026-10-09
- WebMCP Debate: LLM Agents Skip Ads, Putting Ad-Based Business Models at Risk — ConnectRub4818 · 2026-10-09
- At first Grok Bot Meetup, audience-voted idea becomes a 139-marker map app in ~10 minutes — pswider · 2026-10-09
- Amazon AGI Lab: Why 80% Reliable Agents Still Break in Production — AI Engineer · 2026-10-09