# Improvement Agent: Your AI Team Learns From Your Corrections > The Improvement Agent is the AI agent feedback loop in the Ai1 platform by MyZone AI: each week it reads the corrections and feedback your AI agents received, turns repeated ones into clear instruction improvements with client details removed, and applies each change only after you approve it. Canonical page: https://myzone.ai/pages/agents/ai-improvement-agent Part of Ai1, by MyZone AI. Book an Improvement walkthrough: https://calendly.com/d/ct6h-tcy-8qf/ai1-demo?a1=Improvement%20Agent&utm_source=myzone.ai&utm_medium=agent-page&utm_content=ai-improvement-agent-final Last updated: 2026-10-05. Reviewed by the MyZone AI team. Your AI agents learn from your corrections, with your approval on each change. ## At a glance - **Reviews:** What your agents did, the corrections they got and your team's feedback, every week - **Sorts each lesson:** A behaviour change, a business fact or a preference for one agent - **Client details:** Stripped out of every rule, then checked again - **Requests in Slack:** Picked up straight away, not held for the weekly review - **On request:** A ranked audit of one agent, without changing anything - **What it never does:** Apply a change, or approve its own proposal, without you ## What the Improvement Agent does - **Reviews feedback across your AI team:** Each week it reads what your agents have done, the corrections they received and the feedback your team left, so useful lessons are not lost. - **Turns feedback into clear proposals:** Groups similar corrections together and writes each one up as a single, concrete improvement. - **Sorts each lesson to the right home:** Decides whether a lesson is a behaviour change for an agent's instructions, a business fact for your shared knowledge or a small preference for one agent. - **Keeps client details out of instructions:** Rewrites every behaviour change as a general rule with no client names, project names or specific figures, then checks it again. - **Handles requests from your team straight away:** When someone asks in Slack for an agent to work differently, it picks that up at once instead of waiting for the weekly review. - **Audits a single agent on request:** Reviews one agent's instructions, conversations, hand-offs and corrections and ranks what would make it safer, faster or cheaper, without changing anything. ## How the Improvement Agent differs from a feedback log A feedback log collects corrections and nobody reads it. The Improvement Agent reads it for you each week, groups repeated corrections, labels each lesson as a behaviour change, a business fact or a preference and rewrites behaviour changes as general rules with client details removed. You get one plan, approve each change yourself, and every approved lesson goes only to the agents it fits. How this differs from the Continuous Improvement Agent (https://myzone.ai/pages/agents/ai-continuous-improvement-agent): the Continuous Improvement Agent tests changes to your business processes, while the Improvement Agent audits and improves your AI agents themselves and only recommends, with every change yours to approve. ## How it works 1. **Read** (Every week): Goes through each agent's recent activity, corrections and feedback since the last review. 2. **Distil and sort** (After reading): Groups the signal into distinct improvements and labels each as a behaviour change, a business fact or a preference. 3. **Clean up** (Before you see it): Rewrites behaviour changes as general rules with client details removed, and checks them before moving on. 4. **Your approval** (you approve) (Before anything changes): You get one combined plan to review and approve or decline, change by change. 5. **Apply** (Once you approve): Each approved change is passed to the right place: the agent's instructions, your shared knowledge or that agent's memory. ## When to use it - **An agent keeps making the same kind of mistake:** It spots the pattern in the corrections and proposes a rule that fixes it, for you to approve. - **You want one agent to handle something differently:** Ask in Slack. It confirms the plan with you and applies the change once you agree. - **You want your agents to keep improving without managing it:** It reviews feedback every week and brings you a single plan to accept or decline. - **Staff corrections are piling up with no effect:** It turns that backlog into specific instruction updates, so the corrections finally count. ## What you get - A weekly combined plan of proposed agent improvements - Each improvement written as a general rule with client details removed - Every lesson labelled as a behaviour change, a business fact or a preference - A ranked audit of a single agent when you ask for one - Approved changes passed to the right agent or place ## Example: agent improvement audit Example with a fictional company. Names, people and figures are invented to show the agent's output. Any resemblance to a real company or person is unintended. Industry facts are real and cited with their sources. What it was asked: Our tenant enquiry agent has been live for three months: tell me what would make it safer, faster and cheaper, ranked, without changing anything yet. Read-only audit, run the way Make It Better audits any agent: instructions, workflows and skills read in full; all 1,860 conversations counted and 120 read in full; every handoff record, staff correction and cost record traced. Findings are graded Verified, Corroborated or Inferred and scored Impact + Risk + Confidence minus Effort; the rent payment system and email settings were not reviewed, and nothing was changed. Run date: 30 September 2026. ### What it found - The agent sends arrears reminder emails with no person checking them: 64 went out in 30 days, and in 2 of 120 chats read, text a tenant pasted changed the wording. - About half the handoffs to staff were avoidable: 214 of 612 happened only because the agent cannot see the repairs calendar, and 88 more because the answer was missing from its handbook. - Staff left 41 corrections in the feedback inbox, but none reached the agent's instructions, so 17 of them repeat the same three mistakes. A weekly review with a test per fix closes the loop. ### Sources - [LLM01:2025 Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/), OWASP Gen AI Security Project (accessed 2026-09-30) - [LLM06:2025 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/), OWASP Gen AI Security Project (accessed 2026-09-30) ## Guardrails - Nothing is applied automatically. Every proposed instruction change needs your approval, and it never approves its own proposals. - Client names, project names and specific figures are stripped out before any lesson goes near an agent's instructions, and a second check runs before it moves on. - Changes stay within your own Ai1 team. Nothing is shared outside it. - A lesson is only sent to the agents it applies to. ## Frequently asked questions ### How do you improve AI agents from feedback? Collect the corrections your team gives, group the repeated ones and turn each into a clear rule in the agent's instructions. The Improvement Agent in Ai1 by MyZone AI does this weekly, strips out client details and brings you one plan to approve change by change. That is an AI agent feedback loop: the corrections and comments your team gives an agent are collected, turned into clear improvements and fed back into how that agent works. ### Who approves changes to an AI agent's instructions? You do. The Improvement Agent proposes each instruction update in a combined plan, never approves its own proposals and passes an approved change only to the agents it applies to. A fix for one agent changes all of them only if it truly applies to all of them, and even then it comes to you for approval first. ### How often does the Improvement Agent review feedback, and could client details leak into instructions? It does a full review once a week and brings you one combined plan. Requests your team sends in Slack are handled straight away rather than waiting for the weekly review. Client details do not end up in instructions. Each behaviour change is rewritten as a general rule without names, project details or specific figures, and checked again before it is sent anywhere. ### How is the Improvement Agent different from the Continuous Improvement Agent? The Continuous Improvement Agent runs capped, reversible experiments to test what works better. The Improvement Agent learns from the corrections and feedback your agents already received and turns them into instruction updates, one approval at a time. ## More on this topic - [The architecture behind the acceleration](https://myzone.ai/pages/blog/most-productive-day-part3): How a multi-lane human-in-the-loop routing system lets AI agents run autonomously while keeping humans in control of what matters. ## About Ai1 Ai1 is the AI operations platform by MyZone AI, where each client runs on its own private server. The Improvement Agent is not sold on its own: every Ai1 agent, including this one, is included on Developer Pro and every Fully Managed option, with no per-agent charge. Developer Core includes the development agents. Pricing: https://myzone.ai/pages/services/ai1-pricing. Security: https://myzone.ai/pages/security.