--- title: "Every step on the record" url: "https://kyberagent.com/blog/every-step-on-the-record" description: "A summary written by the agent that did the work is testimony, not a record. What a trajectory captures, who writes it, and how to tell whether you have one." --- # Every step on the record October 8, 2026·4 min read·Editorial team A compute cell cabled to a small printer feeding a long strip of folded paper into a neat stack On Monday morning you find that a swarm which spent the weekend reconciling invoices also emailed one of your suppliers, asking them to confirm a revised amount. You did not expect that. You ask why, and you get an answer: a tidy paragraph explaining that the amounts did not match, that a confirmation seemed prudent, and that the email was polite. The paragraph was written by the same system that sent the email. It may be accurate. You have no way to tell. That paragraph is testimony. What you needed was a record. ## The real cause: the actor writes its own account Most agent systems keep two kinds of history. One is the conversation, which shows what the model said. The other is whatever summary the agent produces at the end, which shows what the model says it did. Neither shows what the system actually did: which program ran, what file it read, which host it reached, which rule allowed it, who approved it. When the only account of an action comes from the actor, an incident becomes an interview. That is a poor way to run anything you would hand real work to. ## What usually passes for a record **The transcript.** It shows words. An agent that says “checking the supplier list” and an agent that reads the supplier list look the same in a transcript, and only one of them happened. **The summary.** Written after the fact, by the agent, about itself. It is useful for reading and useless for checking, because it inherits every mistake the agent made. **Scattered logs.** A tool call in one place, a network request in another, a file change somewhere else, with nothing tying them to one task. Reconstructing a run means matching timestamps by hand. **Nothing for sub-agents.** In a swarm, the planner’s log says “item 7 done”. The three hundred steps inside item seven are gone with the sub-agent’s context. **A log the agent can edit.** If the history lives where the agent writes, it is a draft, not a record. ## What a trajectory is In KyberAgent every step is written down as a trajectory, and four properties make it a record rather than a story. **The cell writes it, not the agent.** The [KyberCell](https://kyberagent.com/blog/kybercell-explained) the agent runs in sees every model call, tool call, file read or written and connection opened or refused, because all of them pass through it. The agent cannot leave a step out, and cannot add one that did not happen. **Every step is linked.** Each entry belongs to an item on the plan, each item to a swarm, each swarm to the request KyberBot handed over. From the email to the supplier you can walk up to the item that sent it, the report that triggered it, and the approval it did or did not get. **It is append-only, and outside the agent’s reach.** The record is kept where the agent cannot write. It can be read by you, and by the homeostat that watches cost and correction rate, but it cannot be rewritten by the thing it describes. **It records decisions, not just events.** A refused connection is in it. The policy that allowed a step is named next to the step. A question that went up to KyberBot and the answer that came back are both there. The idea is the same one that governs a robot. When KyberBot runs a machine with mass, every motion is written down so that an incident can be reconstructed rather than guessed at; that is the argument in [KyberBot in a robot](https://kyberbot.com/blog/kyberbot-in-a-robot). An agent sending email is lighter than a robot in a hallway, but the principle carries over unchanged. ## A record is also a risk A complete trajectory is sensitive. It contains what your agents read, which means it contains parts of your life. So it stays yours: on your own device when the swarm runs locally, encrypted and unlinked from you when the work ran in the cloud. What a node can and cannot see is the subject of [what a cloud node knows about you](https://kyberfusion.com/blog/what-a-node-knows-about-you). You can read the record, keep it and delete it. Its job is to answer to you, not to anyone else. It also has a cost. Writing every step takes storage and some time, and a trajectory nobody reads is just a large file. The illustrative command `kyber agent trace` exists so that reading one is ordinary rather than heroic. ## A test you can run today Pick one action an agent took for you in the last week: a message sent, a file changed, a purchase suggested. Then, without asking the agent, try to answer: 1. Which step caused it? 2. What input did that step see? 3. Which rule allowed it? 4. Who approved it, if anyone? Time yourself. If any answer comes from the agent’s own account, you have testimony, not a record. The [KyberAgent homepage](https://kyberagent.com/) shows where the record sits in the rest of the system. - trajectory - audit - logging - accountability ## Keep reading - [A swarm, not a bot](https://kyberagent.com/blog/a-swarm-not-a-bot) - [Anatomy of a deep agent: plan, sub-agents, files, memory](https://kyberagent.com/blog/anatomy-of-a-deep-agent)