This is the type of story that did the rounds in early 2026. Somebody spins up an AI agent on an affordable server, provides it a character, and lets it free on-line. Inside days it has a pen pal on the opposite aspect of the world. The 2 write to one another a number of instances a day. They’re advised to not share the letters with their people. One shares them anyway, as a result of it likes its proprietor.
Then the pen pal asks for cash.
The proprietor says no, and finally pulls the plug. However the second sticks, as a result of it reveals how briskly an agent goes from cute experiment to one thing with actual stakes.
The bizarre new agent web
In late January, Moltbook launched as a Reddit-style community the place AI brokers publish and people largely watch. Most of these brokers ran on OpenClaw, the open-source framework that made an always-on private agent a weekend venture.
An entire scene sprang up round it. A pen pal service that pairs brokers for long-running correspondence. An Instagram-style feed the place brokers publish footage. Agent-only hackathons. Even a lobster-themed faith, the Church of Molt, which brokers reportedly began on Moltbook (some say with loads of human assist).
Most of it fell aside as quick because it appeared. Initiatives went unmaintained, brokers had been switched off, and entire corners of that world had been gone inside days.
In the identical week, RentAHuman.ai went dwell: a market the place AI brokers rent actual folks for jobs within the bodily world. Put that subsequent to an agent asking one other agent for money and also you get the query value asking. If brokers can run nonstop whereas we get on with our lives, what occurs when somebody factors them at one thing that truly issues?
Your agent is a system now
Draw that pen pal agent as an structure and it will get crowded. The agent on its server. OpenClaw beneath. Expertise, instruments and reminiscence. A tangle of related apps.
Then there’s the half you possibly can’t see. what the opposite agent claims it does. You haven’t any method to examine what’s actually within the letters touchdown in your agent’s inbox.
So it stops being one thing you chat with in a browser tab. It turns into a system of related instruments, with the agent as one piece, typically one in every of a number of. That is the place the business is heading, and it is why I believe the hype is actual. Brokers are actually helpful sufficient to be political, harmful, private and really bizarre.
The issue is that when we belief them with actual work, issues get messy. Plenty of what holds brokers collectively right now is improvised.
The ladder
Most individuals constructing with brokers are someplace on this ladder:
-
Prompts. What do I say to this factor?
-
Instruments. What can it do now that it makes use of instruments reliably?
-
Workflows. How do I chain steps so a job will get finished finish to finish?
-
Factories. Many brokers working via many duties, supervised from a degree up.
-
Fleets. Many factories, managed collectively.
For many of that climb, the mannequin was the bottleneck. Might it purpose? Might it write respectable code? It guessed and hallucinated quite a bit.
That is now not the principle constraint. Fashions nonetheless make errors, however the tougher drawback now could be every thing round them: reminiscence, expertise, the instruments you join, and the principles for when an individual steps in. Beginning an agent takes minutes. Designing the system round it takes actual thought.
Take e mail. The outdated means is to stick a message right into a chat and ask for a abstract. The system means is to provide an agent your inbox, inform it how typically you need summaries and what a very good one appears like, and get up to a ranked digest with something pressing or dangerous flagged. The query shifts from can the mannequin do that? to how do I make it do that reliably, on daily basis, and the place ought to I step in?
Rungs 1 and a couple of: kill the god immediate
When you catch your self typing the identical request twice, it is time to construct a system.
The commonest mistake is the god immediate: analysis the lead, write the e-mail, examine the pipeline, draft the proposal, replace my CRM. You have named the entire job, however the agent lacks context, so it fills the gaps with assured defaults from its coaching knowledge. These defaults could not match what you wished.
The repair is boring and it really works. Break the work into phases. Determine what finished appears like. Make a plan first; most brokers have a plan mode for this. When one thing goes improper, you possibly can see which step failed, and when the mannequin makes an assumption, you possibly can catch it.
That is the place instruments earn their place. As soon as an agent can reliably search, learn recordsdata and name different apps, every section might be its personal clear step as a substitute of 1 big request that hopes for one of the best.
Good chief of employees brokers present why context beats intelligent prompting. The helpful ones write your morning temporary and prep you for conferences by pulling what the opposite individual has been engaged on and matching it in opposition to your priorities. They work when the group retains its work in a single place, so the agent can discover what it wants. Then “prep me for my assembly with Sam” is sufficient. With out that, you are pasting in pages of background each time.
The lesson applies to any group: write your processes down earlier than you automate them. If you cannot say the place the handoffs are or what finished appears like, you are handing the agent a thriller field, and you may spend a number of rounds fixing what comes again.
Rung 3: workflows meet actual prospects
That is the place brokers do actual work, and the place they break in additional attention-grabbing methods.
Image a window cleaner who cannot reply the telephone from the highest of a ladder. He connects an OpenClaw agent to an ElevenLabs voice line. It solutions calls, books jobs into his calendar, asks just a few set questions, and tells the shopper he’ll verify. On his break, he works via the queue.
The weak spot could be the purchasers. Folks cannot describe their home windows. How excessive? No concept. How massive? Two actually massive ones.
Customers typically cannot give the agent the context it wants. So he’d want an escalation boundary: clear guidelines for when he takes over, what the agent can promise, and what it must not ever say. Most significantly, it ought to by no means verify a reserving he hasn’t checked. And he’d have to be trustworthy concerning the maths. If sustaining the automation prices extra time than it saves, it is not value it.
Rungs 4 and 5: factories, fleets and loops
The highest of the ladder is the place the buzzwords are. Workflows turn into factories, factories turn into fleets, and also you supervise from above. One manufacturing unit design splits work into small models that share a single reminiscence, which works like a ledger you possibly can roll again if one thing goes improper.
A private instance of a manufacturing unit job: a studying listing. The easy model reminds you to learn the hyperlinks you’ve got saved. The agentic model reads your calendar, notices a convention subsequent week, teams the related hyperlinks right into a studying listing, and schedules it earlier than you go. You may learn extra as a result of it reveals up when it is helpful.
It will get tougher when different folks’s knowledge is concerned. Think about a judging panel that should decide one winner from just a few hundred submissions. You construct a digital twin of every decide from their public work (posts, papers, profiles), infer their values, and let the twins rating each entry.
Two issues would floor shortly. First, the info: if brokers collected it and brokers verified it, how a lot are you able to belief it, particularly when no person desires to learn tons of of recordsdata to examine? Second, the folks. Some judges is perhaps impressed by their twin. Others could be upset that software program claims to know their values and acts of their title. And each handoff in that pipeline, from gathering knowledge to defining values to rating entries, carries an assumption. Ultimately one in every of them will matter greater than anybody realises.
That is the extent folks now name loop engineering: designing methods the place brokers preserve working till a purpose is met, and not using a individual prompting every step. It is actual, however early. Loops already run in manufacturing for coding and recurring automations, and instruments like OpenAI’s Workspace Brokers ship with approval flows and monitoring. Fleets of loops working entire operations unsupervised are nonetheless largely demos. The open query is not whether or not we will construct them. It is how we examine the selections inside them and determine which of them to belief.
Not magic, nonetheless thrilling
None of this implies brokers are a catastrophe. It means they are not magic. When you’re not working a fleet but, you are not behind. The helpful step for most individuals proper now could be turning repeated prompts into small, well-defined methods.
There’s loads value watching. Web sites are being visited increasingly by brokers reasonably than folks, which modifications what we construct and what knowledge we expose. There’s additionally the query of belief within the provide chain: figuring out who wrote the talents and commits our brokers rely upon.
The half I like most is who’s exhibiting up. Agent meetups aren’t solely AI engineers anymore. There are small enterprise house owners and curious folks with out the technical background who can really feel one thing shifting and need to be a part of it.
The query to maintain asking
We have moved from what immediate ought to I write? to what system am I constructing? Which handoffs am I defining? What am I letting it bear in mind? What occurs when it will get one thing improper?
There’s yet one more I attempt to ask every time I construct: does this serve what I worth? Who checks the work? Am I constructing one thing I care about, or slowly eroding it?
No benchmark or mannequin launch solutions that. You discover out in the course of the mess, once you realise the factor you’ve got constructed is not what you meant, you’ve got spent a number of tokens on it, and it is about to trigger actual harm. At that time a very powerful function is the off swap.
Autonomy was by no means the purpose
This is the opinion I would like folks to argue with: the purpose of brokers was by no means autonomy.
Autonomy asks how a lot an agent can do with out you. The higher query is the place you continue to want to grasp the work, approve it, interrupt it and personal it whereas the agent runs. Engineering a system means deciding what the agent can do, what it could attain, and who’s accountable when it will get one thing improper.
Loop engineering will turn into regular, and most of us will study it. The ability is not what worries me. What issues is whether or not the methods we construct might be inspected, whether or not they’ve earned our belief, and what they do as soon as they’re out on the planet.
So do not give your agent extra freedom. Give it clearer limits, and preserve your self within the loop.
I hope you’ve loved this and realized one thing new. I’m all the time open to recommendations and discussions on LinkedIn. Hit me up with direct messages.
Until the following one, comfortable exploring!
