This is the most common misunderstanding we meet, and it is an entirely reasonable one. You open a model, you type, it answers well, and it feels like you have an assistant. What you have is a conversation. An agent is a different kind of object, and once you can see the parts, the difference stops being marketing and becomes obvious.
It waits for you. It knows what you told it in this conversation and forgets it when the window closes. It has no duties, no schedule, and no idea what it did yesterday.
It boots into an identity, loads what it learned before, runs a cycle whether or not you are watching, and refuses its own bad output before it reaches you.
Not a personality. A contract. Who it is, what it owns, what it is never allowed to do, and which rules outrank its own judgement. This file is read at the start of every session, so a rule written once survives every restart.
A chat window has a prompt you paste in and hope you remembered. An agent has a contract it cannot start without.
# boot sequence, read in this order 1. MEMORY what I learned that outlived the session 2. STATE what is true right now, and what is held 3. SOUL identity and execution discipline 4. INBOX directives waiting for me 5. HEARTBEAT the cycle I run without being asked
A skill is a written procedure the agent loads when the work calls for it: when to use it, what it must check, what it must refuse. The value is not that the model is capable. It is that the right procedure arrives at the right moment without anyone remembering to ask.
The difference between a model that could do something if prompted perfectly and a system that does it the same way every time.
name: visual-qa-screenshot use when: any page is about to ship runs: 24 checks across 3 viewports verdict: pass or fail, per check on fail: blocks the deploy
Every cycle ends by writing what happened and what was learned, including the mistakes. The next session reads it. That is how a correction made on a Tuesday is still in force in November, and how the same wrong turn stops being taken twice.
Most people mean chat history when they say memory. Chat history is a transcript. This is a decision, made about what deserves to persist.
2026-08-02, banked An availability check that only imports a library reports the tool as present, then reads every result as clean. Rule taken: run the thing, and treat any failure as unavailable, never as clean.
This is the part almost nobody has, and it is the part that decides whether a system can be left alone. A warning depends on somebody reading it. A refusal does not. Ours block the action: spending stops at a ceiling, a deploy stops on a failed check, an image containing text is rejected before it can reach a file.
Ask of any safeguard you are shown: does it stop the thing, or does it print a message and continue? Only one of those works while you are asleep.
PASS a check ran and found nothing wrong BLOCK a check ran and found the problem QUEUE nothing could check it, so a person looks before it ships the third verdict is the important one: unavailable never reads as clean
The agent wakes on a schedule, decides what most needs doing, does it, and records what it produced. Nobody assigns the task. If it has nothing to do it says so plainly, and if it is blocked it names what it is blocked on rather than going quiet.
This is the whole difference between something you use and something that runs. You are not in the loop by default. You are in it when you choose to be.
every cycle writes three records presence so the system knows it is alive status so a person can read it trail so tomorrow inherits today writing one obligates the other two, and the cycle reports itself INCOMPLETE until all three are current
Every part above is a decision someone made once and wrote down so it holds forever. The boot order, the procedure that loads at the right moment, what gets remembered, what gets refused, what happens at three in the morning with nobody watching. A model supplies the intelligence. None of it supplies the structure, and the structure is the product.
That is what we build. One operation, scoped in writing before anything starts, delivered as a working agent you own and run on your own subscription. The agents on this site are the same thing, built for our own business, which is the only reference we think is worth much.
Request a scoped build See the agents we runExamples on this page are taken from the system we operate daily and simplified for reading, with internal identifiers removed. Nothing here is illustrative invention.