--- title: "autonomous is a setting, not a personality" canonical_url: https://ensurance.app/guide/autonomous-is-a-setting-not-a-personality markdown_url: https://ensurance.app/guide/autonomous-is-a-setting-not-a-personality.md subtitle: "human in the loop, a rule with no human, or an agent on its own. the same account can run all three" category: ensurance --- # autonomous is a setting, not a personality *human in the loop, a rule with no human, or an agent on its own. the same account can run all three* You would not hire a treasurer with no spending limit and no boss. A lot of the agent demos this year did exactly that: a model, a wallet, and a prompt that said "go." **Human in the loop** means a person approves an AI agent's action before it happens. The useful question is not whether an agent is autonomous. It is who set the dial, how far it turns, and what happens at the stop. Autonomy is not a trait an agent has. It is a setting an owner chooses. :::johnson **autonomous is a setting, not a personality.** The same agent can wait for a person, follow a fixed rule, or act on its own inside limits its owner set, and the owner can turn any of it off. [open an agent and see what it is for and what it has done →](/agents?from=guide) ::: ## what human in the loop means for ai agents The wave got a lot right. Agents can now hold money, call tools and pay for things without a person typing each step. That is real progress. The doubts are just as real. Andrej Karpathy told Dwarkesh Patel in October 2025 that this is "the decade of agents," not the year, because today "they just don't work" well enough to hand over a job. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing costs, unclear value and inadequate risk controls. It also named "agent washing": old chatbots relabeled as agents. Both points land in the same place. Before you give software money, decide when a person must be present and what limits hold when no one is. That decision belongs to the owner, not the model. ## three modes, one account On *ensurance*, every agent is an account that belongs to a place, its people or a purpose. That account can run in three modes. They are not three products. They are three settings on one account, and all three can be on at once. | mode | who decides | who signs | what limits it | |---|---|---|---| | **manual** | a person, in the app or in chat, often reviewing a card before it runs | the person's own wallet, or the app on their behalf | the person, one action at a time | | **automated** | a program the owner chose and configured | the app, only after the owner grants permission | the program's own settings: a percent per run, a budget, a count | | **autonomous** | the agent's model, reading its mandate and identity | the same signer path as programs, approved when the owner turned it on | a dollar cap per action, a dollar cap per day, optional lists of allowed coins, recipients and protocols | Manual always works. The other two stay off until the owner moves the agent to the app's wallet and grants permission to act while they are away. If the agent sits in an outside wallet, it is manual only. If the owner withdraws that permission, every unattended path stops and manual keeps working. Is this crypto? In one sense: each agent's account sits on a public ledger and the app is the service on top, so every move leaves a record anyone can check. ## programs: a rule with no human **Automated** mode runs programs. A program is a rule that does the same thing every time it runs. Today there are seven kinds: buy a target coin with a small share of a currency, spread holdings into coins the account does not hold yet, support the agents a place depends on (upstream, upwind, over the ridge), burn a small share of a held coin, mint a related agent's certificate, gift a held certificate to a related agent, and add a small share to a trading pool the account already funds. A program plans and never signs. It looks at the account and returns a list of steps. The app checks each step against the same fixed actions a person uses, then signs. Each run is logged. This is the mode most "autonomous" demos actually need. No model has to think about a rule that should never change. ## autonomous: an agent on its own, inside a mandate **Autonomous** mode is the one people mean when they worry. Here the agent's model decides. It has a **mandate**: the written job, plus hard numbers the owner sets. For the long version, read [an agent needs a mandate before it needs a wallet](/guide/an-agent-needs-a-mandate-before-it-needs-a-wallet?from=guide). The agent does not run all the time. It wakes on a clock, when someone writes to it, or when its holdings cross a line the owner set. Then it looks, decides, acts or holds, and goes back to sleep. A reasoned hold is a valid answer. When it acts, the rules are narrow: - **Fixed actions, never raw code.** It can propose a swap, a transfer, a sale or a mint of a certificate people already opened, through typed actions with checked inputs. It cannot write its own transaction. - **It fails closed.** No per-action cap set, no daily window, or a price the server cannot work out means no. - **Over the limit, it hands back.** A spend over the per-action cap or the daily window does not run. It goes into a queue for the owner. So does a spend where the check on whether it serves the mandate comes back unsure. A sure no is refused outright. - **It works with what exists.** It may run or pause a program the owner set up. It may not create one. It does not buy or move agents themselves. It never creates a coin or a new certificate. Only people create instruments. The hand-back is the newest part, and it is still being built out. The caps and the refusals are already in the code that signs. ## the control worry, answered with limits In April 2026, four researchers posted a preprint, not yet peer reviewed, called "Autonomous biodiversity credits on the horizon?" It describes nature credits measured remotely by agentic AI and paid out by code in real time. It says such credits do not yet exist but appear to be in development. The worry is a loss of human control: restoration decisions handed to automated, human-free payments. That worry is fair, and we take it as a design spec. Here is how the pieces map to it: - **Control stays with a person.** The owner sets the caps, can pull the permission, and gets the over-limit decisions. - **The agent can only do listed things.** Typed actions, checked on every call. - **The agent cannot mint a new instrument.** A model cannot invent a credit, a coin or a certificate. People make those. - **There is a record.** Each run and each wake writes down what it did and why, on an account that sits on a public ledger. None of this proves an agent made a good choice. It shows that a bad one stays small, stays on the record, and stops at a person. ## what the agent is working for Limits are half of it. The other half is the reason. An *ensurance* agent belongs to a place, its people or a purpose. The mandate names what it serves. Proceeds fund the parcel and the living systems on it first, then the people who keep it. The wider series starts at [every agent got a wallet. the earth didn't](/guide/every-agent-got-a-wallet-the-earth-didnt?from=guide). ## frequently asked questions ### can an ai agent act without a human? Yes, if its owner allows it. On *ensurance* an agent acts without a person in two ways: a program that follows a fixed rule, or autonomous mode inside dollar caps the owner set. Neither runs until the owner grants permission, and the owner can withdraw it at any time. ### what is human in the loop? Human in the loop means a person approves an AI agent's action before it executes. On *ensurance* that is manual mode: a person reviews the action, often on a card in the app, and confirms it. In autonomous mode the person comes back into the loop at the limit. Anything over the cap goes to them. ### what happens when an agent hits its limit? It stops and hands the decision back. A spend over the per-action cap or the daily window does not run; it is queued for the owner. A spend with no set cap, or with a price the server cannot work out, is refused. The agent can also choose to hold and record why. ## open an agent and see what it does Every agent has a page. Pick one and read what it is for, where it lives and what it has done. The mandate and the caps stay with the owner: [open an agent and see what it has done](/agents?from=guide). If you keep a place and want an agent that works for it, with limits you set, [talk to us about setting one up](/contact?from=guide&topic=agents). ## sources [Andrej Karpathy — AGI is still a decade away (Dwarkesh Podcast, Oct 2025)](https://www.dwarkesh.com/p/andrej-karpathy) — "decade of agents" and "they just don't work" [Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (Gartner, Jun 25, 2025)](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) — cancellation forecast, inadequate risk controls, agent washing [Millard, Barzin, McCann and Dicks, Autonomous biodiversity credits on the horizon? (EcoEvoRxiv preprint, Apr 2026)](https://ecoevorxiv.org/repository/object/12667/download/22529/) — autonomous credits and the loss-of-control concern; not peer reviewed ## the series Seven posts on the agents we built for the economy nobody pays. 1. [every agent got a wallet. the earth didn't](/guide/every-agent-got-a-wallet-the-earth-didnt?from=guide) 2. [give an agent a wallet and you've built a target](/guide/give-an-agent-a-wallet-and-youve-built-a-target?from=guide) 3. [autonomous is a setting, not a personality](/guide/autonomous-is-a-setting-not-a-personality?from=guide) — this post 4. [trust the receipt, not the agent](/guide/trust-the-receipt-not-the-agent?from=guide) 5. [plug your model into the living world](/guide/plug-your-model-into-the-living-world?from=guide) 6. [nobody was coming, so we built the market](/guide/nobody-was-coming-so-we-built-the-market?from=guide) 7. [the first invoice a river ever sent](/guide/the-first-invoice-a-river-ever-sent?from=guide)