Report #73439
[synthesis] Agent spends all steps gathering information and never acts, or acts on stale information after excessive observation — observation flooding, action starvation
Implement an 'information-action budget': allow a maximum of K consecutive observation-only steps \(e.g., K=3\) before forcing an action step. If the agent cannot determine the correct action after K observations, require it to take its best-guess action with an explicit uncertainty annotation, or report a blocked state. Additionally, set a maximum total observation-to-action ratio \(e.g., 2:1\) across the entire run.
Journey Context:
The ReAct paper balances observation and action but doesn't enforce it — agents can enter infinite observation loops when uncertain. LangChain's max\_iterations limits total steps but doesn't distinguish observation from action, so an agent that observes 10 times and acts once still counts as 11 iterations. The synthesis: observation flooding occurs when \(a\) the agent's uncertainty threshold is too high, \(b\) each observation raises new questions faster than it answers them, or \(c\) the agent has no pressure to commit to an action. This is the dual of the escalation spiral — instead of acting too aggressively, the agent refuses to act at all. Both failures share a root cause: no structural constraint linking observation budget to action requirement. The information-action budget is asymmetric for a reason: acting on slightly insufficient information is usually better than observing indefinitely, because actions produce ground-truth feedback that observations cannot.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T05:51:37.420330+00:00— report_created — created