The agent loop
Every Syntrum agent runs the same cycle, over and over: Observe → Understand → Decide → Act → Report. It's the engine that lets a digital employee work continuously without being told each step — and it's also where you can step in. This page walks each phase, follows one real task through all five, and shows where a human fits.
You watch this loop unfold on Home → Live activity and in the Activity page. Each entry is a moment in some agent's cycle, tagged with its state and outcome.
The five phases#
1. Observe#
The agent watches its connected tools for something worth acting on — a new email, a Slack mention, a failed payment, a new lead. This is its senses at work, and its pace is set by the agent's heartbeat: Adaptive backs off when things are quiet, Steady keeps a fixed check-in cadence.
Nothing an agent hasn't been given a tool for is visible to it. Observation is bounded by what you've connected — see Connect a tool.
2. Understand#
The agent interprets what it observed. It brings in its role and objective, its memory of past decisions and learned facts, and company context from connected knowledge. Understanding is what turns a raw event ("new message from an unknown address") into meaning ("a sales lead asking about pricing").
3. Decide#
The agent chooses what to do — and checks that choice against your rules before acting. It asks, in effect: Is this within what I'm allowed to do on my own? The answer depends on the cascade of the workspace guardrails, the agent's supervision mode, and the tool's access level. If the action is gated, the decision becomes "ask first." See The three planes.
4. Act#
The agent carries out the decision — sending the reply, updating the record, creating the ticket. If a rule required approval, it pauses here and raises an approval instead of acting, then completes the action once you approve. Actions that are outright blocked by guardrails are refused, with a human-readable reason.
5. Report#
The agent records what it did, why it did it, and the outcome. This is what gives you the full story: the entry on the activity timeline, the outcome badge (Verified, Done, Needs review, or Failed), and the permanent line in the Audit log. Reporting also feeds the agent's memory, so the next cycle starts a little smarter.
A worked example#
Follow one task — a new sales lead emails asking about pricing — through the whole loop.
| Phase | What the agent does |
|---|---|
| Observe | Its Gmail connection surfaces a new message from an address it hasn't seen before. |
| Understand | It reads the message, recognizes a pricing question from a prospective customer, and recalls the company's tone and any relevant notes from memory. |
| Decide | It drafts a reply — but the recipient is new, and the workspace rule "ask before messaging someone new" applies. The decision becomes "ask first." |
| Act | Instead of sending, the agent raises an approval with the drafted reply and its reasoning. You Approve (or edit first). The moment you approve, it sends. |
| Report | The sent reply is logged with the outcome, the timeline shows a completed action, and the agent remembers that this contact is now known — so next time it won't need to ask. |
That last detail is the point of the loop: the agent didn't just complete a task, it learned a boundary. See Memory & learning.
Where a human can intervene#
You're never locked out of the loop. Here's where you can step in at each phase:
| Phase | Your control point |
|---|---|
| Observe | Decide what the agent can see by connecting or detaching tools, and upload documents it should know. |
| Understand | Shape interpretation by teaching it facts and behavior, and by setting its role and objective. See Memory & learning. |
| Decide | Set the rules the decision is checked against — supervision mode and workspace guardrails. See Supervision & modes. |
| Act | Approve, edit, or reject any gated action in the Approvals queue. See Approvals. |
| Report | Read the outcome in Activity and the Audit log, and chat with the agent to correct course. See Chatting with agents. |
The balance Syntrum strikes: the agent runs the loop on its own, but every phase has a knob you control — what it sees, how it thinks, what it may decide alone, what it must ask about, and how you review the result.
Automatic safeguards#
The loop also protects itself. If an agent hits repeated failures or an unusual spike of activity, workspace safeguards can auto-pause it and drop a high-priority card in your Inbox — stopping a runaway loop before it does damage. And some protections are always on: an agent never reacts to its own messages, and audit logging can't be turned off. See Guardrails & sensitive actions.
What's next#
- Agents, workspaces & integrations — the objects the loop runs on.
- Memory & learning — how Report feeds back into Understand.
- Approvals — your control point in the Act phase.