Home / How it works

What happens after you press Enter.

Every request, whether it's a question, a Flow, a BRD or an integration, goes through the same stages. The model is one step inside them. The steps below are what make the output safe to deploy.

The command pipeline

Eight steps, in order.

01

Remove secrets

secret redaction

Before a command is saved or sent to a model, Stripe, Slack, AWS and GitHub keys, tokens and key=… pairs are removed. A pasted key comes back as [REDACTED_STRIPE_KEY].

02

Classify the request

intent classifier

Rules pick the action (read, answer, analyze, create, update, delete, deploy) and the scope (field, Flow, Apex, security, integration, data) without an AI call. Object names are masked first, so Delete_Log__c can't trigger a delete. Only ambiguous text goes to the model.

03

Use the conversation

thread memory

A follow-up with no subject, like "which are they", carries over the previous turn's intent. Conversations are stored per user and come back when you reload.

04

Plan and rate the risk

router

The request becomes a plan of steps with a risk level. Reads run straight away. Writes get generate → review → validate → approve → deploy → verify. "Delete" with "all", or any mention of production, is rated critical automatically.

05

Check against your org

grounding + Org Brain

The plan is checked against your real metadata and Org Brain before generation, so standard objects are reused and existing fields aren't duplicated.

06

Build, check, retry

self-healing loop

Nova generates the change, checks its structure (Flows and Apex both), and dry-runs it against the org. A real org error is fed back and it tries again, up to three times. If it still fails, it stops and tells you why instead of shipping something broken.

07

You approve

the gate

You see what will change, where, at what risk, and how to undo it. By the time Approve appears, Salesforce has already confirmed the deploy would succeed. Nothing is written before this step.

08

Deploy and verify

deployment + read-back

Deploys go to a sandbox first, and the org-type matrix decides what's allowed where. Afterwards Nova reads the metadata back. A Flow can deploy cleanly and still be broken, so a mismatch triggers a rollback and a retry.

From document to deployment

How a BRD becomes a plan you approve.

Ten steps with two loops: one checks the reading for gaps, the other asks you questions. A person always approves before anything is deployed.

01

Receive

The upload returns right away and reading runs in the background. Uploading the same file again reuses the existing session.

02

Guard

Page and size limits and a kill switch run before any AI cost. Anything stuck for more than 10 minutes is marked failed.

03

Read

Returns one structured result: summary, objects and fields by area, future items, open questions, and a citation for each requirement.

04

Check the read

A gap loop re-reads sections that came back thin before moving on.

05

Ground

Every proposed object and field is compared with the org. Standard objects are used first.

06

Account

The coverage banner: every requirement is either mapped or set aside with a reason.

07

Clarify

Concrete options with a recommended default. You answer once, and it doesn't guess.

08

Plan → approve → deploy

Objects, fields, validation rules, permission sets, record types and Flows. Send the plan for review, approve it, then deploy through the loop above.

The agent

For longer tasks, an agent with a budget.

Multi-step work runs as an agent loop. It has a step and cost budget, clear reasons for stopping, and a fixed set of tools, including running Apex tests, reading a Flow and searching Org Brain.

Budgets

Each run has a step and cost limit. When it hits one it stops and reports why, rather than continuing.

Policy before tools

Every tool call is checked against policy first. Production targets, external tools and instructions found inside data it read are flagged.

Ask once

You can set approval so the agent asks you, once per run, before production changes. Your workspace admin decides how much it can do on its own.

See the pipeline run

Watch a build on a real org.

We'll run a request against a sandbox and show you each step as it happens.