Field Guide · anatomy
How AI Agents Actually Work: The Loop, in Front of Your Team
An AI agent runs a loop — perceives, plans, acts, reports. The interesting part of the loop is the report, which posts to your team's Slack so the work is visible to everyone. This is what makes an agent different from a model.
The phrase “AI agent” hides a specific shape of computation that buyers have an easier time evaluating once they have seen it laid out. An agent is a model running a loop. The loop has four parts. The first three are interior to the agent — perceive, plan, act. The fourth — report — is what makes the loop legible to the team that hires the agent. Most explanations of agent loops dwell on the first three parts because they are the technical achievement. The fourth is what the buyer is actually paying for. This piece walks through the loop in the order it runs, and lingers on the report step because that is where the agent stops being an experiment and starts being a teammate.
Why it matters
An agent without a visible report step is, from the team's perspective, a black box that does work and the team finds out later. That shape is uncomfortable for a buying decision because the buyer cannot watch the loop run. A loop with the report step in Slack — every action the agent takes summarized in a channel the team already reads — collapses the black box. The buyer can see the perception that triggered the action, the plan that was chosen, the action that was taken, and the result. The loop is the same in both cases; the difference is whether the team can read it.
The agent loop has four parts. They run on a clock — every few minutes for active agents, every few seconds for high-frequency ones, on a calendar trigger for scheduled ones.
1. Perceive.
The agent reads the inputs it has been wired to. For a customer service agent, that's new tickets in the queue and recent posts in the support Slack channel. For a marketing agent, that's competitor mentions and dashboard alerts. The perceive step is bounded by the constitution; the agent does not perceive things outside its stream. A request that arrives in a channel the agent does not read does not exist to the agent.
2. Plan.
Given what it perceived, the agent decides what to do. The plan step is where most of the model's intelligence lives — the agent considers options, reasons about which actions match its posture, checks against its boundaries, and produces a sequence of actions. If two actions conflict, the agent applies its conflict-resolution rules. If no permitted action fits, the agent escalates rather than improvises.
A useful way to think about the plan step is that it is a junior version of what a teammate would do — read the situation, propose an approach, check it against what's allowed, and either proceed or ask. The plan step is fast in absolute terms (seconds, sometimes a minute or two) but it is the part of the loop where the agent's posture and constitution actually do their work.
3. Act.
The agent executes the plan. Drafting the brief, posting the rollup, updating the wiki, escalating the ticket, opening a thread to ask a clarifying question. The act step is bounded by tools — the agent can only do what its tool surface permits. If the plan calls for an action the agent does not have the tool for, the loop ends with a flag in the team's review queue, not with the agent improvising a workaround.
4. Report.
The agent posts what it did, why, and what it didn't do, in the team's Slack. This is the step that buyers undervalue and the platform considers load-bearing. The report names the input that triggered the loop, the plan the agent chose, the action it took, the result. The team reads the report the way they would read a teammate's standup update — quickly, in passing, with the option to dive into the thread for the reasoning if anything looks off.
The report is what turns the loop from a black box into a teammate's work. A loop without the report step still runs — the agent still acts — but the team's relationship to the agent's work is mediated by reports the team has to go find. With the report posted automatically, the team's relationship is direct. The agent's work is in front of the team the same way a coworker's work is.
What runs between loops.
Between iterations, nothing happens. The agent is waiting for a new input. This is the part of agent design buyers most often misread — they imagine the agent thinking continuously between loops. It does not. The agent is reactive between loops the way a person at a desk is reactive between tasks: the agent isn't doing anything when nothing has arrived. Each loop is its own discrete unit of work, with a beginning, a plan, an action, and a report. The loops compound across days — Tuesday's report shapes Wednesday's perception — but each loop ends cleanly.
The edge
Almost every meaningful difference between AI agents in 2026 happens at the report step. The perceive, plan, and act steps are getting better across every vendor as the underlying models improve. The report step is a design choice. Vendors who put the report in a private dashboard build agents that are private to the manager who hired them. Vendors who put the report in the team's Slack build agents the team treats as teammates. The technical work in the loop is similar; the deployment pattern at the report step is what determines whether the loop is a feature or a relationship.
Honest take
This four-step loop is a working model, not the most precise one. Real agents have sub-loops, parallel branches, and tool-use sequences that complicate the picture. The plan step in particular is sometimes two or three rounds of model calls, with intermediate perceive steps in between. What is described here is the shape that holds at the level a buyer needs to read — close enough to true that asking about each step gets you a useful answer, simplified enough that the buyer can hold all four parts in their head while they read a Roster entry. The honest version of the loop acknowledges that under the hood the picture is messier; the legible version of the loop matches what the buyer can see in Slack.
An agent runs a loop. Perceive, plan, act, report. The loop is what makes an agent different from a model; the report is what makes the loop different from a black box. Most of what a buyer should care about happens in the report.