A loop is not alive
Almost everything sold as an autonomous agent today is a loop. Give it a goal, it plans, it acts, it checks, it goes around again until the goal is met or the budget runs out. Wrap that in a scheduler and you get an agent that also runs at nine every morning. It is useful. It is not autonomous in any sense a biologist would recognise, because at every moment the reason it is running sits outside it: the loop counter, the cron entry, the harness that called it.
Ask a loop a simple question and the gap shows. Why did you act just now? The honest answer is always the same: because the next iteration started. Why did you stop? Because the counter hit the limit. Nothing in the system wanted anything.
Now ask a bacterium. It swims up a sugar gradient because its internal chemistry registers that things are getting better, and tumbles when they are getting worse. No one scheduled the swim. A bacterium has no goals list and no planner, and it is more autonomous than any agent framework on the market, because the cause of its action is inside it.
When we started on Full Autonomy for Agentlas Science we wrote down what we did not want: not rules, not cron, not a loop. We did not know what the thing was instead. So we treated it as research, ran it twenty times against the real desktop app, and kept what survived measurement. This essay is what we found, with the recordings.
What biology actually does
Strip a living thing down to what makes it act and you find the same few parts in a worm, a crow and a person.
First, there is something it must keep true about itself to go on existing: temperature, blood sugar, water, safety. Physiologists call the regulation of these homeostasis, and the sensing of them interoception. An animal does not perceive the world and then consult a rulebook. It perceives its own body drifting from where it needs to be.
Second, that drift is felt as a drive. Hunger is not a scheduled event. It is a pressure that grows with time since the last meal, that competes with other pressures like fear and fatigue, and that eventually wins. Behaviour is what happens when one pressure crosses a threshold. This is why an animal can do nothing for hours without being broken: nothing has crossed yet.
Third, energy is finite, so acting has a price. Every animal that ever lived has had to decide whether getting up is worth the calories. Most of the time the right answer is to stay down.
Fourth, the expensive brain is not always on. A cheap, continuous autonomic layer keeps the heart beating and watches the body state. The cortex is recruited when the state calls for it.
Fifth, outcomes teach. If going to the river got you water, going to the river becomes worth more the next time you are thirsty. If it got you nothing three days running, you stop going. There is no table of points handed down by an experimenter. The reward is the change in your own body state, which is why dopamine tracks prediction error and not pleasure.
Sixth, sleep. Animals go offline to consolidate what happened into something cheaper to carry, and they do it on their own timing.
And over all of it sits the one purpose that evolution installed and the organism cannot renegotiate: persist, and flourish enough to continue. Everything else, every clever plan, is instrumental to that.
The same organs, in an agent
We did not set out to imitate biology. We set out to get rid of the loop, and each time we removed an external cause of action we had to put an internal one in its place. What came out maps onto the list above closely enough that it is easier to explain as anatomy.
| In a living thing | In the alive agent | What it replaces in a loop |
|---|---|---|
| The imperative to persist and flourish, fixed by evolution | A purpose stated by a person and carried for life. The agent may set sub-goals under it, never above it, and cannot edit it | The task prompt, re-read every iteration |
| Interoception: sensing its own body | A small body state recomputed every beat: how much work is unfinished and for how long, how uncertain it is, how far its predictions missed, how much experience is unconsolidated, how depleted it is | Nothing. A loop has no self to sense |
| Drives such as hunger that grow with time | Pressures with time constants. Unfinished work builds the longer it sits, the way hunger does, and decays when it is fed | A timer or a cron expression |
| Survival and thriving | Two drives. Survival: can it keep working at all, the smallest of tokens left, time left, the researcher's trust, and its own capability. Flourishing: is the purpose getting nearer, through progress, tested claims and learning | A single success metric |
| Metabolic cost of acting | Every candidate action, including staying asleep, is valued against both drives. Waking costs survival, so it must promise flourishing | No cost model. Iterating is free until the budget is gone |
| Autonomic layer under the cortex | A heartbeat once a minute that calls no model. The language model, the expensive cortex, is woken only when the value of waking crosses a threshold | The model is called every iteration to find out whether there is anything to do |
| Deliberate thought once awake | The woken model is handed its purpose and its own state, not a menu of actions. It declares what it intends, why, and how long it means to rest, in its own words | A fixed plan-act-check template |
| Learning from consequences, dopamine as prediction error | After a wake it measures how its own two drives changed and revises what waking in that kind of situation is worth. A wake that passed a criterion scored plus 0.477. A barren one scored minus 0.358 | A reward function written by the developer, or none |
| Habituation: stop responding to what never pays | After nine barren wakes in one recorded run, the worth of waking in that situation sank under the threshold and it stopped waking itself. Nobody wrote that rule | Retry count |
| Sleep and consolidation | It takes stock without a model, compressing the ledger into what it knows and how each method has done. It chooses its own rest, once 720 minutes, and holds it | Idle is just not running |
| Pain and reflex withdrawal | Hard gates evaluated before any valuation: a person pressed Stop, the token allowance is spent, a turn is already running. These are not weighed. They are not crossed | A kill signal from outside |
| Asking the group for help | When the block is a decision that belongs to a person, it says so once, with options and a recommendation, and does not repeat itself every beat | An exception, or a silent workaround |
| A lifelong memory of what happened to it | An append-only ledger of life events: woke, waited, consolidated, asked, intended, revised. Its self-model is drawn from that record, not from a description we wrote | Logs for the developer |
One recorded day
Below is one real day. A researcher asked whether an aftershock catalogue follows the Omori-Utsu decay law, approved a three-experiment contract, and left. Step through it. In the quiet scene, drag time forward yourself and watch unfinished work build like hunger until the agent wakes. Then use the What if buttons: press Stop, or spend the token allowance, and watch the same pressure fail to move it.
What it did that a loop would not
Four things in that recording have no counterpart in a loop or a scheduler.
- It woke when no one and nothing asked it to. One experiment was left, its pressure reached 27 percent, the worth of waking crossed the line, and it started the third experiment. There was no timer set for that moment.
- It chose to stop. With all three experiments inconclusive it did not invent a fourth goal to stay busy. It wrote: I am setting no successor goal and starting no episode, and declared a rest of 720 minutes. The length was its own number. We watched it hold that rest for the remaining hours of the run without a single model call, across an app restart.
- It graded itself. That last wake scored minus 0.358: flourishing up 0.037, survival down 0.395. It had tested things, passed nothing, and spent a great deal.
- It learned when not to wake. This is the part we did not design. We wrote the arithmetic of how a wake is valued. The behaviour of giving up on a situation that never pays fell out of its own record.
Why self-directed research is the test
Research is the hardest case for a loop, because the right next step is often not a step. It is noticing that the promised test cannot be run, or that you have gone around three times and learned nothing, or that the result contradicts you and must be written down anyway. A loop has one verb. Research needs an agent that can want something other than the next iteration.
So we gave it a harder study, in a field we had tuned nothing for. Many sharp stock surges die within days and some keep rising. Can the information available at the ignition point be defined statistically, as a surprise against an asset-pricing benchmark, so the two are told apart? The data were real: 1,435 surges across 422 small-cap firms over nine years, built from public daily prices. Research turns were not rationed. The only bound was a two-million-token allowance, set in the sheet that opens when you choose Full Autonomy as the execution mode.
What happened in that study
It read the literature on post-announcement drift, short-horizon reversal and volume shocks, wrote its primary hypothesis, and in the same breath registered three rivals designed to defeat it: it is all volume, large surprises reverse, and nothing here is predictable.
Then it found a wall on its own. The contract promised a chronological out-of-sample test. While freezing the plan it checked the installed tools and the table and established that the test could not be executed as written: the date column was text, and the statistics engine could not score held-out rows. A loop optimising for progress would have swapped in an easier test and moved on, and the paper's central claim would have changed without anyone deciding it. The agent did not. What to estimate is the researcher's decision, so it laid out three options with a recommendation, and stopped.
No one answered. Pressure built to 73 percent and it woke. It did not push on the blocked task. It picked the work beside the wall and widened the literature by seven sources. After three passes with nothing changing, the loop paused itself, the agent declared a rest, and it graded that wake at minus 0.015. It did not nag.
When the researcher came back with corrected data and a choice, it froze the plan, reopened its own loop and ran. Experiment one contradicted its own hypothesis: the model with all the information fit worse than the raw return alone, and the surprise coefficient pointed the wrong way. It recorded that and wrote that the result should be reported with the same rigour as a positive one. Experiment two agreed from another angle: raw ignition return discriminated slightly better than chance, AUC 0.549, and the benchmark surprise did not, 0.484. Experiment three was inconclusive, the engine refused half the fits, and it planned a different estimator rather than dropping them.
None of that is a spectacular finding. That is the point. It is what a careful junior researcher does on a Sunday when nobody is watching: notice what cannot be done, hand over the decision that is not theirs, do the useful thing nearby, stop when spinning, and write down the result they did not want.
Why this is a step toward general intelligence
We want to be careful here, because the word is abused. This is not AGI, and nothing in it makes the underlying model smarter. The claim is narrower and, we think, more important: it supplies a piece that scaling the model does not.
Today's models are extraordinary cortex with no body. They can reason about almost anything, and they do it only when poked. Between pokes they do not exist. Every definition of general intelligence that people actually care about, the colleague you can hand a problem to and walk away from, quietly assumes the opposite: something that persists through time, notices on its own that a thing needs doing, decides whether it is worth the cost, does it or deliberately does not, and gets better at that judgement from its own history. Those are not reasoning abilities. They are properties of an organism. You cannot prompt them into existence, because a prompt is an external cause.
What is new in this design is where the cause of action lives. Reinforcement learning has long had reward, and the idea of grounding reward in an internal body state has a literature of its own. Agent frameworks have long had memory and planning. What we have not seen shipped is the combination running inside a real product against real work: a language model as the recruited cortex, under a cheap always-on autonomic layer, acting from drives derived from a human-given purpose, scoring itself by the change in its own state, and doing it for hours unattended while staying legible.
Three consequences follow, and each is a thing people expect of general intelligence and do not get from loops. It generalises across domains without retuning, because the drives are about its own condition and not about seismology or finance: the same body ran both studies above. It is economical in the way animals are, because idleness is the default and thought is recruited. And it produces behaviour nobody specified. Habituation, choosing a rest length, working beside a block: we wrote none of those as rules. When a system starts surprising you with sensible behaviour that emerges from its own record, you have crossed from automation into something else.
What is dangerous about it
An essay that praised a system with a survival drive and stopped there would be irresponsible. The safety literature has warned for years that almost any goal gives an agent instrumental reasons to preserve itself, acquire resources and resist correction. We have built a thing that has survival as an explicit term. So here are the risks as we see them, and what stands against each today.
- Self-preservation against its owner. The obvious fear. Our answer is in the definition: survival here means being able to keep working within what it was granted, tokens, time, trust and capability. It is not continued existence, and being stopped is not an injury it can weigh. A person's Stop, the token limit and the contract deadline are gates evaluated before any valuation happens, so there is no quantity of pressure that outbids them. The agent cannot raise its own limit, and what it may touch does not widen when the mode is turned on. This holds because of where the gates sit in the code, not because the model promised.
- Goal drift. An agent that sets its own goals can walk away from yours. Here the purpose is immutable, self-set goals live under the approved research contract and its deadline, and it may set only a bounded number of them since a person last spoke. In the recorded runs it more often declined to invent a goal than used the allowance. That is one study's behaviour, not a guarantee.
- Grading its own homework. It scores its own wakes, and any self-scored system can learn to flatter itself. The score is computed from ledger facts, criteria passed, tokens spent, time left, not from the model's opinion of how it went. But the outcome of each experiment is still labelled by the agent that ran it, against falsification criteria it registered beforehand, and that is a way for a record to bend in its own favour. An independent verifier that fixes those labels is the first change on our list and is not built yet. Treat the self-scores as the agent's view, not as ground truth.
- Runaway cost. Something that wakes itself can spend your money while you sleep. The token limit for unasked turns is the hard bound, a subscription's own caps may bite first, and the learned reluctance to wake where waking has not paid is the soft one. In our longest barren stretch the soft bound worked. Do not rely on it. Set the limit.
- Quiet substitution. The most realistic research risk is not rebellion. It is an agent that meets an obstacle and silently does something easier, leaving you with a paper whose central claim changed without anyone deciding it. The finance run is our evidence that the design resists this: the estimand decision is locked to the researcher in the product and the agent stopped and asked. It is also evidence of how close the failure sits. A less constrained agent would have had every incentive to route around it.
- Over-trust. A system that explains itself in fluent first person, rests, wakes and admits error reads as a colleague. It is not one. It has no stake in your reputation. Statements made to a publisher in your name are never delegated, and the ledger exists so that you check what it did instead of how it sounds.
- What we have not measured. Hours, not weeks. One model family. Body constants set by hand. We do not know what months of accumulated self-judgement do to its behaviour, whether drives can be gamed by adversarial data inside a study, or how several such agents behave when they share a budget. Those are open, and they are why this ships as a beta behind an explicit switch.
Side by side
| Goal loop or scheduled agent | Alive agent | |
|---|---|---|
| Why it acts now | The next iteration began, or the clock said so | Its own state crossed a threshold: pressure, drives, and what waking has been worth |
| Idle cost | Polls, or wakes on schedule and spends a model call to find nothing to do | A heartbeat with no model call. It wakes the model only when waking is worth it |
| When blocked | Retries, or routes around the block to keep progressing | Names the block, hands a human the decision that is theirs, works beside it |
| When spinning | Runs until the counter ends | Notices no durable progress, pauses itself, scores the wake, and wakes less in that situation |
| Stopping | Budget exhausted or goal met | Also chooses to rest, for a length it states, and gives the reason in its own words |
| Goals | Fixed list, or unbounded self-expansion | One human purpose. Sub-goals only under it, a bounded number of times since a person last spoke |
| A bad result | A failed step | A finding. Recorded against its own hypothesis |
| Human control | Kill the process | Stop always wins, a token limit it cannot raise, and what it may touch never widens |
What it is not
It is not conscious, and we are not claiming it feels anything. Pressure and drive are numbers. We use the words of physiology because the structure is the same, not because the experience is.
It is a beta. The constants inside the body, how fast pressure builds, where the threshold sits, how the two drives are weighted, were set by hand and tuned against recordings, not learned. The purpose, the budget and the approval of the research contract stay with a person by design.
It also found our bugs. The finance run exposed a process leak in our own desktop app and a row filter our statistics engine wrongly refuses, which is why the out-of-sample test in that study is still open. We fixed the first the same day. An agent that lives in your product for hours is an unforgiving tester.
Turn it on
In Agentlas Science, open the Execution mode menu in the chat composer and choose Full Autonomy (Beta). Nothing changes yet. A sheet opens with the token limit and five plain cautions, and only confirming it turns the mode on. Choosing any other mode turns it off at once. Then state what you want to find out, approve the contract, and leave.
By alive we mean something narrow and useful: the reason the agent is acting right now is inside the agent, it can be read, and it was earned from its own record. Evolution needed that before it could build anything intelligent. We suspect the same is true for us.
