Why agent stacks can keep taking valid steps while losing the verified state
Present-Quantity-813 · reddit · 2026-07-24
When agent systems loop, correctness depends on state, not just the next action
The post argues that agent frameworks like ReAct, reflection, and graph orchestration do not by themselves guarantee reliability. A system can keep taking locally valid actions while having lost track of the verified world state or goal position.
Core idea
- The author frames reliability as:
- P(correct step) = P(correct position) × P(correct entrance | position)
- “Position” refers to the current goal, evidence, authority, and remaining obligation.
- “Entrance” refers to the next tool call, transition, or action.
Main warning
- A graph can constrain allowed transitions, but it does not prove the agent is still in the right state.
- Termination conditions can become stale when the world changes.
- In one system the author watched, the agent entered a 3-day recursive loop, generated 102 work items, and 68% were repair tasks.
Open question
The post asks where termination really lives in an agent stack:
- the prompt
- a graph node
- a supervisor
- a budget
- or an externally verified state transition
It also asks what prevents old evidence from authorizing another iteration after the goal or world state has already moved on.
More from coding & agent
- A LangGraph learning journey through agents, blackboards, and bottlenecks — Clear-Dimension-6890 · 2026-07-24
- New ComfyUI tile-refine node automates high-quality image upscaling with context anchoring — blakeem · 2026-07-24
- Codex adds GPT Live voice control for fully hands-free coding on desktop — op7418 · 2026-07-24
- T3 Code nightly defaults to DeepSeek reasoning traces unless set to null — pvncher · 2026-07-24
- AREX introduces a recursively self-improving deep research agent with inner and outer loops — _reachsumit · 2026-07-24
- A local open-source agent skill flags the rest of your GitHub diff — Saboo_Shubham_ · 2026-07-24