# How it works

The Datapace loop in five steps: connect within an agreed scope, infer, confirm with your experts, agents propose, your team approves and records.

## The loop

Every job in Datapace, from documenting an estate to right-sizing an instance, follows the same five steps. The first three build the graph; the last two put it to work.

![The Datapace loop: connect, infer, confirm, propose, approve and record, with the graph in the middle](/docs/how-it-works-v1.svg)

## 1. Connect within an agreed scope

Which databases and environments Datapace may read is agreed with your team before anything connects. Datapace is designed to read within that scope only: schema metadata, workload signals, and sample structures. It does not need a copy of production.

## 2. Infer

From what it reads, Datapace infers what the data means: entities, measures, relationships, and lineage. Each inference carries a confidence score. The infrastructure facts around each database (instance class, storage, cost, performance, usage, freshness) land on the same graph. See [the graph](/docs/concepts/the-graph).

## 3. Confirm

Your experts accept or reject each inference in a few working sessions. Nothing enters the confirmed graph unvalidated, which is what makes it documentation people and AI systems can trust. See [infer and confirm](/docs/concepts/infer-and-confirm).

## 4. Propose

On the confirmed graph, the agents draft the routine work: the FinOps agent proposes right-sizing and flags idle databases, the Performance agent proposes index and rewrite work and overdue upgrades, the Quality agent proposes checks, the Migration agent proposes field-level mappings, the Documentation agent keeps the docs current. Every proposal carries its evidence. See [agents](/docs/concepts/agents).

## 5. Approve and record

Every action goes through policy. The risky ones wait for a person. Whatever happens, by an agent or a person, is written to the audit ledger with who, what, on which database, the evidence, the approver, and the outcome. See [policy, approval, and audit](/docs/concepts/policy-approval-audit).

## What stays constant

- **Agents propose, people approve.** No proposal executes without your team's approval.
- **The graph is the product.** Every agent, every AI system, and every dashboard works from the same confirmed graph.
- **Governed context, never raw access.** Copilots, BI, and MCP clients reach production data through the graph and its gate. See [governed context](/docs/concepts/governed-context).
- **Across engines.** The loop is the same for every engine you run. See [across engines](/docs/concepts/across-engines).