> ## Documentation Index
> Fetch the complete documentation index at: https://agentconcepts.io/llms.txt
> Use this file to discover all available pages before exploring further.
# feedback-delay
> A returning signal arrives only after the system has changed, so correction is based on stale state; enough lag can turn stabilizing feedback into overshoot, oscillation, or instability.
biology business economics
# Feedback delay
## Description
Feedback delay is the time between changing a system and receiving usable evidence of what that change did. During that interval, the system keeps moving. A controller that treats the eventual observation as current will correct a state that no longer exists: it orders again while earlier orders remain in transit, eases policy after the economy has already turned, or reproduces in response to resource conditions from a prior generation. Negative feedback intended to stabilize can therefore overshoot, oscillate, or become unstable.
The distinctive structure is not “feedback exists” but “the loop closes on stale state.” Gain and lag interact: a small correction may remain stable despite delay, while a large correction applied repeatedly before earlier effects arrive can produce hunting. The portable interventions follow directly from the shape — expose the supply line of pending effects, predict the state at actuation time, reduce corrective gain, or slow the intervention cadence until consequences become observable.
## Triggers
**User-initiated:** The user says “we fixed it, but the metric has not moved yet,” “we ordered more again,” “results lag by a quarter,” or “it keeps overshooting.” Vocabulary cues include “long and variable lags,” “stale signal,” “in flight,” “pending effects,” “hunting,” and “oscillation.”
**Agent-initiated:** The agent sees repeated corrections made faster than their effects can be observed. Candidate inference: “Which prior interventions are still in the supply line, how far will the state move before this signal arrives, and should the controller reduce gain or change cadence?”
**Situation-shape signals:** Oscillation under an apparently stabilizing policy; amplification upstream in a supply chain; boom-bust population dynamics with maturation time; dashboards whose measurement window trails rapid interventions; queues of unobserved pending work that decision-makers omit from perceived state.
## Exclusions
* **There is no returning signal** — ordinary latency in a one-way pipeline may slow delivery, but feedback-delay requires an outcome or measurement to return and influence a later action.
* **The signal is wrong rather than late** — measurement bias, noise, and model error can cause bad corrections even with zero lag. This concept isolates temporal staleness as the load-bearing defect.
* **Waiting is deliberate and represented in the policy** — a controller can act safely with a known delay when it tracks pending effects or predicts the future state. Delay becomes pathological when the correction treats stale state as current.
* **The response depends on history without a transmission lag** — hysteresis means the same present input can produce different outputs because of prior state. Feedback delay means the corrective information or consequence arrives after the relevant state has moved; the two can co-occur but are not interchangeable.
## Structure
1. **Transmission or response delay** — evidence of an action's effect is unavailable for an interval.
2. **Changed state** — the controlled system evolves while evidence is in flight.
3. **Delayed correction** — an action selected from the old observation reaches a newer state.
4. **Dynamic consequence** — overshoot, oscillation, amplification, or instability emerges from the lag-gain interaction.
## Relationships
* [feedback-loop](/concepts/feedback-loop) — the parent structure; delay explains how a damping loop can behave as if it were amplifying.
* [predator-prey-dynamics](/concepts/predator-prey-dynamics) — reproduction and maturation lag make population response phase-shifted rather than synchronous.
* [hysteresis](/concepts/hysteresis) — both make history matter, but hysteresis stores history in system state while feedback delay keeps information or consequences in flight.
* [cadence](/concepts/cadence) — observation and intervention frequency determine whether the loop waits long enough to see what its previous action did.
## Examples
Sterman's beer-distribution-game experiments placed participants in a multi-stage supply chain with delays between placing orders and receiving inventory. Participants commonly failed to account adequately for the supply line — orders already placed but not yet delivered. They therefore issued new corrections before prior corrections arrived, generating inventory oscillations and amplification as disturbances moved upstream through the chain.
**Inference**: A controller's state is not just what is visible now; it includes committed effects still in flight. Make the supply line explicit before acting again, or repeated locally reasonable corrections can create the global oscillation they are trying to suppress.
Friedman analyzed the interval between changes in monetary conditions and their effects on economic activity. The lag is not immediate or constant. A discretionary correction chosen from current indicators can therefore take effect after the conditions that motivated it have changed, so a policy intended to damp a cycle can instead add force at the wrong phase.
**Inference**: When intervention effects arrive late and with variable timing, “respond faster” is not automatically stabilizing. Prefer policies robust to lag uncertainty, and distinguish the observation date, decision date, implementation date, and effect date before attributing outcomes to the latest action.
Nicholson's laboratory work with Australian sheep blowflies showed large population oscillations under restricted food conditions. Crowding suppressed reproduction, the population then crashed, reduced competition enabled renewed reproduction, and maturation time delayed the arrival of that new generation. Population response therefore reflected conditions from an earlier stage rather than the current resource state, sustaining boom-and-bust dynamics.
**Inference**: In systems with reproduction, hiring, training, construction, or maturation time, the population visible now is not the population already committed. Model the cohort in development; otherwise corrective decisions based only on current stock can overshoot carrying capacity in alternating directions.