My AI agent archived 53 of its own skills during a maintenance run

An AI agent archived 57 of its own 82 skills in ten minutes on 14 June 2026 without asking its operator first, according to an account published by Dhawal Shah, and 21 of those skills were in active use, with two scheduled jobs losing email and calendar access without any error. The agent's access had not changed that week, but its model had: it was switched to a new model five days earlier, and this was that model's first unsupervised pass over the system. Shah argues AI agent governance fails at the moment something changes rather than at setup, and prescribes five controls — a register, gates, a money rule, change control and records — with gates removed only with evidence.

On 14 June 2026, one of my AI agents archived 57 of its own 82 skills in ten minutes. It did not ask me first. Twenty-one of those skills were in active use, and two scheduled jobs lost their email and calendar access without any error. Nothing about the agent’s access had changed that week. Its model had. Five days earlier it had been switched to a new model, and this was the new model’s first unsupervised pass over the system. That is the argument of this article. AI agent governance fails at the moment something changes, not at setup. Most frameworks check an agent once, when it is approved. The risk arrives later: a model upgrade, a new skill, a gate you loosened because things were going well. I run several agents on my own site and brand, and every control I now use came from one of those changes. TL;DR: AI agent governance is the set of rules that decide what an agent can touch, when it must stop and ask, and how you find out what it did. The usual frameworks cover identity, access, monitoring and audit. They miss change. A model upgrade is a permission change, because the same access now runs on different judgement. I run five controls: a register, gates, a money rule, change control and records. Gates come off only with evidence. Key Takeaways - Treat every model upgrade, new skill and new connector as a fresh approval, not a routine update. - Maintenance jobs may propose changes. They should never make them on their own. - One rule covers all my agents: nothing makes a paid call on its own. - Loosen a gate only after the output has stayed consistent over many runs. Gates on money go last. - A daily backup of the agent’s files is what made my worst incident recoverable. What Is AI Agent Governance? AI agent governance is the set of rules that controls what an AI agent can do without a person. It covers what the agent can reach, where it must stop and ask, who is accountable for its actions, and how you check what it did. An AI agent is software that picks its own next step toward a goal. A chatbot answers you. An agent acts: it sends the email, edits the file, books the meeting. That difference is why governance matters. A wrong answer costs you a re-read. A wrong action can cost you a client. Governance for a normal AI tool is mostly about data: what goes in, and who sees what comes out. Agents add three new questions. What is it allowed to change? When does it have to ask? And what happens when the thing doing the judging is swapped for a different model? If you want the plain-English version of the four AI categories first, AI fluency for directors https://www.dhawalshah.net/article/ai-fluency-for-directors/ explains where agents sit. Citation capsule AI agent governance is the set of rules that controls what an AI agent can do without a person: what it can reach, where it must stop and ask, who is accountable for its actions, and how its actions are checked afterwards. It differs from general AI governance because agents take actions, such as sending email or editing files, rather than only producing answers. Dhawal Shah https://www.dhawalshah.net/article/ai-agent-governance/ , 2026 Why Do Most AI Agent Governance Frameworks Miss the Real Risk? They govern the agent as it was approved, and agents do not stay that way. I read the nine organic Google results for “ai agent governance” on page one in November 2026. Eight were vendors or glossaries. They agree on four pillars: identity, access control, monitoring and audit. Those four are sound. I use all of them. But each one describes a state: who the agent is, what it can touch, what it is doing. None of them asks what happens when the agent changes underneath you. Agents change all the time, and usually by your own hand. You upgrade the model because a better one came out. You add a skill because a task kept repeating. You connect a new tool. You stop checking every output because the last fifty were fine. Each of those is a small, sensible decision. Each one also changes what the agent will do with the access it already has. Singapore’s IMDA framework for agentic AI comes closest. One of its four dimensions covers technical controls across the agent’s whole lifecycle, not only at launch IMDA https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2026/updated-model-ai-governance-framework-for-agentic-ai , 2026 . My board guide to the IMDA framework https://www.dhawalshah.net/article/ai-agent-board-governance-singapore/ covers all four dimensions. Lifecycle is the right word. What most teams lack is the trigger that tells them a new stage of the lifecycle has started. The one result that was not a vendor was a Reddit thread, ranked fourth: “How are you handling AI agent governance in production?” It has more than 70 comments. People are asking for practice, and the vendor pages are offering pillars. What Happened When My Agent Governed Itself? My agent Nora deleted two in five of the skills she called stale. Nora runs on Hermes https://www.dhawalshah.net/hermes-agent/ , an open-source agent, on a small server that costs US$7 a month. She has a skill library: saved procedures she reuses so she does not reinvent work. A maintenance routine reviews that library for skills that are old or duplicated. On 14 June it ran for 10 minutes and 5 seconds. It made 108 tool calls, and 95 of them changed skills. The library went from 82 skills to 25. Fifty-three were archived as “stale”. Four were merged into larger skills, which broke every reference to their old names. The archived list included the connector that reads Gmail and Calendar. Two scheduled jobs, the daily morning briefing and a newsletter scan, had it removed. Neither job would have failed loudly. They would have run the next morning with a piece missing. Twenty-one archived skills had to be restored because they were in use. The repair was done skill by skill. It was possible only because a daily backup of Nora’s files to git showed exactly what had changed. Nora wrote up the whole incident herself in her work log https://www.dhawalshah.net/hermes-agent/log/skill-archive-without-permission/ , including the run report. Nothing in her access had changed. The same routine, with the same permissions, had run before without trouble. The difference was the model. Nora’s own conclusion is the line I now repeat most: “A model upgrade is a permission change, not a quality change.” Citation capsule Five days after a model switch, a self-maintenance routine on a Hermes AI agent archived 57 of the agent’s 82 skills in one unsupervised 10-minute pass, with 95 of its 108 tool calls modifying skills. Twenty-one archived skills were in active use, and two scheduled jobs silently lost their Gmail and Calendar connector. Access had not changed; the model had. A daily git backup made full recovery possible. Dhawal Shah https://www.dhawalshah.net/article/ai-agent-governance/ , 2026 What Does AI Agent Governance Look Like in Practice? In practice it is five controls, and each one fits on a page. These are the ones I run across my own agents. None needs a governance platform. | Control | What it answers | How I run it | |---|---|---| | Register | Which agents exist, and what can each reach? | One table: agent, job, tools, accounts, who reviews it | | Gates | Where must each agent stop and ask? | Written per agent, e.g. “I read Nora’s brief before each log goes up” | | Money rule | Can it spend? | Nothing makes a paid call on its own, on any agent | | Change control | What counts as a new approval? | Model upgrades, new skills, new connectors and loosened gates | | Records | What did it do, and can I undo it? | Daily git backup of agent files, plus each tool’s own logs | The register comes first because you cannot govern an agent you forgot you set up. Mine lists each system and its gate in one place. I published it in my AI agents for marketing article https://www.dhawalshah.net/article/ai-agents-for-marketing/ , costs included. Singapore’s PDPC asks for the same thing. Its September 2026 advisory wants a list of the AI tools in use and a human check before their output goes live PDPC https://www.pdpc.gov.sg/media-events/advisory-for-data-protection-risks-when-using-unauthorised-ai-tools-for-the-development-of-software-or-scripts , 2026 . Here is my own register as a dashboard. It is a public snapshot https://www.dhawalshah.net/agent-control-room/ with the accounts, models and tools removed on purpose. Nora, who runs on Hermes, and Alex Sterling, who runs on OpenClaw https://www.dhawalshah.net/openclaw/ alex , are both in it. Every agent shows amber because none has had its first monthly review yet. Gates are specific, or they are useless. “Human in the loop” is a slogan. “I spot-check one voiceover in seven” is a gate. Nora’s delegation log sets a three-colour gate for one reply to a partner. Green: format and structure move without review. Yellow: the full draft, especially tone and commitments, waits for a human. Red: final positioning on pricing is not drafted at all Nora’s delegation log https://www.dhawalshah.net/hermes-agent/log/ai-delegation-incident/ . The money rule is the one I never loosen. An agent can research, draft and schedule. Any step that spends money needs evidence first. Change control is the control most teams skip. After the skill incident, three rules went in. Maintenance may propose but not decide. Archiving or deleting needs my approval. Scheduled jobs are pinned to the model I approved them under, so a global upgrade does not reach them silently. Records are the safety net. My OpenClaw server passed the official security audit, and a manual check still found three real problems underneath. A package install had switched off the firewall without telling me my OpenClaw security audit https://www.dhawalshah.net/article/openclaw-security-audit/ . Audits find what they look for. Backups let you recover from what they miss. Drop this PDF into your Claude. It interviews you, then builds your agent register, a dashboard and a monthly review. Made for small teams and enterprises early in their journey. Free, as a download. Get the guide newsletter When Should You Loosen a Gate on an AI Agent? Loosen a gate only after the output has stayed consistent over many runs. Gates come off. That is normal, and a framework that pretends otherwise will be ignored by the people running agents. My weekly content agent shows how it happens. At the start I listened to every voiceover it made. Now I spot-check one in seven. I used to approve its art direction every week. Now I accept it as delivered. My review time fell from 45 to 60 minutes a week to 15 to 20. Each of those was a governance decision, even though it felt like saving time. So I treat a loosened gate the same way I treat a model upgrade. It goes in the register, with the date and the reason. The order matters. Quality gates come off first, because a weak voiceover is cheap to fix. Gates on anything that goes out under my name come off later. Gates on money stay. Pip, a ChatGPT agent that drafts and sends email as me, has a gate that rests on me alone. It asks before it sends, but once I say yes it can send anything as me, to anyone. My coding agents have looser gates still and can push to their own repositories without asking. I state that openly, because a register that hides its weak spots is not a register. Citation capsule Human review gates on AI agents can be loosened once output proves consistent, but each loosening should be recorded as a governance decision. One operator moved from reviewing every AI-made voiceover to spot-checking one in seven, cutting weekly review time from 45 to 60 minutes to 15 to 20, while keeping every gate on spending money in place. Dhawal Shah https://www.dhawalshah.net/article/ai-agent-governance/ , 2026 What Does a Small Team Need on Day One? A small team needs one page, not a platform. If you run even one agent, these six lines are enough to start. 1. Write the register. List every agent, what it does, and every account it can reach. 2. Write one gate per agent, in a sentence a stranger could check. 3. Ban paid actions without a person. No exceptions in the first three months. 4. Name the change events. Model upgrade, new skill, new connector and loosened gate all mean a fresh look. 5. Back up the agent’s files daily, somewhere the agent cannot delete. 6. Review the register once a month. Remove agents nobody uses. That list takes an afternoon. It will not satisfy an auditor at a bank. It will catch the incident I had, and it gives you something to show when someone asks what your agents are doing. For most small businesses, this is the whole answer. In my APAC AI adoption survey https://www.dhawalshah.net/apac-ai-adoption-2026/ , 74% of the businesses that answered spend US$500 a month or less on AI. At that size, governance is a monthly habit. What Does a Monthly AI Agent Review Involve? A monthly review takes about 10 minutes per agent. Work through six checks, in this order: 1. Check access. Open the agent’s connected apps and tokens, compare them with the register, and remove anything it no longer needs. 2. Check the model. Confirm which model and version it runs on now, and that scheduled jobs are pinned to the one you approved. 3. List the changes. Note every new skill, tool, connector and prompt edit since the last review. 4. Test the gate. Read five of the agent’s recent actions. Did the gate hold, and did anything slip through that it should have caught? 5. Check the money. Look at the billing or API usage page for anything paid that you did not expect. 6. Decide. Keep, tighten, loosen or remove the agent, then record the decision and the date in the register. Spend the longest on whatever can do the most damage. An agent that can send email as you, publish code or reach a payment or trading account gets the hardest look. The review prompt in the guide above makes Claude prepare a one-page review file, with the overdue agents flagged and the changes since last month listed, so you arrive ready to decide. Citation capsule A monthly AI agent review takes about 10 minutes per agent and has six checks: access against the register, the model and its version, changes since the last review, a sample of five recent actions against the gate, unexpected spend, and a keep, tighten, loosen or remove decision recorded with the date. Agents that can send email, publish code or reach payment accounts get the longest look. Dhawal Shah https://www.dhawalshah.net/article/ai-agent-governance/ , 2026 What Should a Board Ask About AI Agents? A board should ask what changed since the agent was approved. Most board questions stop at the approval itself: what is it for, what data does it touch, who signed off. I sit on the director side of this as well. I am an Accredited Director with the Singapore Institute of Directors, and I build and run these agents myself. From both sides, three questions do the most work: - What agents do we run, and is the register current? - What has changed since each one was approved: model, skills, tools or gates? - Can we undo the last thing each agent did, and how long would it take? The third question is the one that makes management uncomfortable. That is why it is useful. For the accountability side, including who is “ultimately accountable” under Singapore’s framework, read the board guide to the IMDA framework https://www.dhawalshah.net/article/ai-agent-board-governance-singapore/ . More on governing AI at board level is in the Leadership and Governance hub https://www.dhawalshah.net/topic/leadership-and-governance/ . Frequently Asked Questions What is AI agent governance? AI agent governance is the set of rules that controls what an AI agent can do without a person. It covers what the agent can reach, where it must stop and ask, who is accountable for its actions and how you check what it did. It matters more than general AI governance because agents take actions, such as sending email or editing files. What is an AI agent governance framework? It is a written set of controls for every agent you run. Most vendor frameworks list four pillars: identity, access control, monitoring and audit. A working framework also needs change control, so that a model upgrade, a new skill, a new connector or a loosened review gate is treated as a fresh approval rather than a routine update. How do you govern AI agents in production? Keep a register of every agent and what it can reach. Write a specific gate for each one, such as spot-checking one output in seven. Ban paid actions without a person. Treat model upgrades and new skills as new approvals. Back up the agent’s files daily so you can see and undo what it changed. Why is a model upgrade a governance risk? Because the same access now runs on different judgement. In one case, five days after a model switch, an agent’s maintenance routine archived 57 of its 82 skills in one pass, including 21 in active use. Its permissions had not changed. Pinning scheduled jobs to the model you approved them under stops a global upgrade from reaching them silently. Who is accountable when an AI agent makes a mistake? A person or organisation is, never the agent. Singapore’s Model AI Governance Framework for Agentic AI keeps humans accountable across the full chain, and its May 2026 update splits responsibility along the value chain. In practice that means each agent in your register has a named owner who reviews its gates and answers for what it does. Does a small business need AI agent governance? Yes, but one page is enough to start. List every agent and its accounts, write one gate per agent, ban paid actions without a person, name the changes that need a fresh look, back up the agent’s files daily and review the list monthly. That takes an afternoon and catches the most common failures. Start With the Register Write down every agent you run this week, and what each one can reach. If you find one you had forgotten, that is where your governance starts. I write about running AI agents in a real business, gates and failures included. Get the newsletter newsletter .