Feedback loop

A worked example

During the annual workplan review at a shared services centre, an internal auditor asks an approved planning agent to sequence proposed engagements for a draft workplan due to the front office by close of business.

Starting from: Proposed engagements, risk ratings, estimated effort, staff availability, required coverage and scheduling constraints.

1. The agent produces a first draft by matching the proposed engagements with available staff and planning periods.
2. A validator checks the draft and reports that a required area is not covered and that one specialist is assigned to overlapping engagements.
3. Those findings are routed back to the agent. It adds the missing coverage, changes the sequence and reallocates the overlapping assignment.
4. The revised draft is checked again. If a constraint still fails, the finding triggers another revision. Once the checks pass, the draft is sent to the internal auditor for substantive review and clearance.

Result: The workflow returns a draft workplan that satisfies the encoded planning constraints, together with a record of the failed checks and revisions. The internal auditor can then assess whether the result is sensible, including whether the agent optimised what was easy to score while overlooking less readily quantified concerns.

Where you may hear it

• An annual workplan review in an internal audit function
• An ICT governance board discussion on controls for autonomous agents
• A vendor demo during a procurement meeting
• A design session for an automated workflow at a shared services centre
• A post-implementation review after an automation pilot

Often confused with

• Observation and action

  Intermediate

  In an acting system, observations are the signals it receives, actions are the steps it takes, and the act-observe loop is the repeated cycle linking the two.

• Reinforcement learning

  Intermediate

  A training method that aims to improve behaviour through trial, feedback and cumulative reward across a sequence of actions.

• AI agent

  Starter

  An AI system that can pursue a goal by selecting and carrying out multiple steps, often using software tools, with oversight matched to its risks.

Common misconceptions

• ✗Common misconception: A feedback loop means the system always gets better each time.

  ✓Correction: Not necessarily. Improvement depends on the quality of the feedback and the comparison rule. If the feedback is noisy, delayed, biased or based on the wrong target, repeated adjustment can make performance worse rather than better.

• ✗Common misconception: Feedback loops are the same as reinforcement learning.

  ✓Correction: They are related but different. Reinforcement learning usually refers to a training approach where a model is updated over many episodes using reward signals. A feedback loop can happen during one live task at run time, even if the model weights never change.

• ✗Common misconception: Any sequence of observe then act counts as a feedback loop.

  ✓Correction: A single observe-act step is not enough. The defining feature is that the observed result affects a later action through repetition or correction. Without that return path, there is no loop.

Quick check

The simple version

It is a repeat-and-check pattern: do something, look at what happened, and use that result to improve the next step.

The technical version

In AI systems, a feedback loop is a structure in which an agent or workflow produces an action and routes an observation or evaluation of its outcome into a later step. The system may compare that signal with a goal, constraint or scoring rule and adjust the next action. Feedback may come from the external environment, a software tool, a validator, a user or another model. The loop may be explicit, such as retry-until-valid, or more complex, such as plan-execute-evaluate-replan. A feedback loop does not by itself imply model training. It can operate entirely at inference time, with no parameter update. Observation and action name the signals and steps in an acting system. A feedback loop emphasises the return path by which an outcome affects later behaviour. Design choices include the feedback source, delay, noise tolerance, stopping condition, human escalation and safeguards against oscillation, runaway retries or optimisation of a poor proxy.

Reviewed external resources

• Primary sourceModel Context Protocol (opens in a new tab)Model Context Protocol

  This link explains the Model Context Protocol and how tools can provide structured information back to a model, which suits technical readers interested in how feedback is passed between systems.

  Deeperlink checked 2026-07-31

• EncyclopaediaFeedback (opens in a new tab)Wikipedia

  This link gives a broad introductory explanation of feedback as a general idea across systems, which suits readers who want the core concept before applying it to AI.

  Intermediatelink checked 2026-07-31

• ArticleBuilding effective agents (opens in a new tab)Anthropic

  This link explains how AI agents take actions and use results from those actions, which suits readers connecting feedback loops to agent behaviour.

  Intermediatelink checked 2026-07-31

• VideoVideo explanations: Feedback loop (opens in a new tab)YouTube search

  This search results page offers multiple video explanations of feedback loops from different creators, which suits readers who prefer to compare visual teaching styles rather than rely on one source.

  search results, not one pagelink checked 2026-07-31

• DiscussionCommunity discussion: Feedback loop (opens in a new tab)Reddit search

  This search results page shows community discussions and examples of how people talk about feedback loops in practice, which suits readers who want informal explanations and debate but should read critically.

  search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.