Skip to the content

Prepare a works-council or DPIA pack

Generate templates that describe what your deployment collects, keeps and sends, and switch the fleet to the non-monitoring profile during talks.

Built, not yet claimed preview: this guide describes a package that has not passed its release gate.

Generate the template pack that describes what your AgentGarm deployment collects, keeps and sends, switch your fleet to the non-monitoring profile so no per-user content reaches Central Lite while an agreement is negotiated, and check what Central Lite can then still see. The templates are not legal advice: they state product facts and leave every legal question to your counsel.

1. Generate the pack on a machine

On any machine with AgentGarm, agentgarm compliance pack --out writes six Markdown files into a directory it creates: README.md, data-inventory.md, processing-description.md, dpia-skeleton.md, non-monitoring-profile.md and dora-supplier-metadata.md. They are generated from the build's own event classification, the retention table, the catalog's dated jurisdiction entries and the compliance profile agentgarmd reports, so they describe what is deployed. An existing file is never replaced: use a fresh directory for each pack.

The pack is Markdown only. A PDF and an export from the console are not built.

For outbound inspection preview, non-monitoring-profile.md has a section for assessing policies that list sites: what it reads, keeps, sends and for how long. It reads the path, query and body of the requests launched agents send to the listed sites, in memory on the machine, never a response, a header or another site. A finding keeps the site, the rule, a count, where in the request and what was done, never the secret or the body. The request's path is kept, redacted, on the machine only, never uploaded, not even as a hash. Central Lite receives findings, uninspected records, alerts and ten-minute counts per site, the site as a keyed hash. Findings are kept as security records and the counts as operational ones, for the retention those classes have. Its placeholders ask which sites are listed and why, and whether the people covered were told.

agentgarm compliance pack --out works-council-pack

Checked by the walk: this command ran as written on a Linux node enrolled in Central Lite.

2. Complete the templates with your counsel

Every document starts with “Template — not legal advice”. The pack states what the product does; everything a law may require is left to you, as an [operator: …] placeholder (legal bases, recipients, purposes) or a [counsel: …] question. Its legal references are only the catalog's entries, such as GDPR Article 88 and BetrVG § 87(1) no. 6, each with its source and the day it was read, as orientation for counsel; whether one applies to you is for you and your counsel to decide.

The DORA supplier metadata carries placeholders and questions only: no clause text ships until counsel has reviewed it. AgentGarm's own counsel review of these templates is still open, so treat them as a first draft for your works council or DPIA, not a finished document.

3. Decide what may leave the machines

The non-monitoring profile is for deployments that must not monitor individuals, for example while a works agreement is negotiated. Under it a machine sends no tool call, model call, file or network event to Central Lite: its activity leaves only as one record per hour with counts per data class, agent kind and policy outcome. Security and audit events are still sent, with every field that identifies a person removed, not pseudonymised. Support bundles are not uploaded, a CI job's ask is not relayed to a remote approver (it is denied), and agentgarmd replay export, which would carry one user's commands and paths to Central Lite's replay, is refused. On a Mac with the computer-use airbag preview, its records keep no element title and the capture monitor's no app name. The airbag records only its decisions on protected targets, never what was typed, and nothing of input that passes. Outbound inspection preview keeps no request path at all, on the machine included, and sends its findings and alerts without the site, and its counts only per listed entry as the admin wrote it, with no task or person.

Enforcement and the machine's own records and audit log do not change. A policy can ask for the profile too, in default.yaml; a default.yaml that does not load applies it, so privacy fails closed.

compliance:
  profile: non_monitoring

4. Switch it on from Central Lite

On the Central Lite host, garm-central compliance show prints the compliance settings Central Lite signs for nodes: the version, the org's retention profile and non-monitoring switch, and each node's own. compliance non-monitoring on asks every node for the profile, or one node with --node; the reason (1 to 256 characters) is chained, and the next signed settings reach the nodes over the control channel, not at once: a connected node gets them on its next delivery, an offline one when it next connects. A node applies the profile when the org's switch, its own switch or its signed policy asks, so switching it off at Central Lite never turns off what a policy asks.

The profile is an operator setting, not a protection against the operator: an admin can switch it off. Every change is chained on the machine, raises a compliance_profile alert at Central Lite, and shows on the machine's compliance profile line in agentgarm status (compliance profile NON-MONITORING once it is in force) and its local Privacy & retention screen, so the people concerned can see it. Until the settings arrive, that line still shows the previous profile: run it again.

garm-central compliance show
garm-central compliance non-monitoring on --reason "works agreement under negotiation"
agentgarm status

Checked by the walk: these commands ran as written on a Central Lite host and a Linux node enrolled in Central Lite, and the output showed compliance profile NON-MONITORING.

See also

Set retention, legal holds and data-subject requests

Set how long Central Lite keeps each class of record, put a legal hold on the records a case needs, and answer a person's access or erasure request.

5. Check what Central Lite can still see

In the console, open Transparency. “What nodes send” shows the fleet's compliance profile (non-monitoring, standard or mixed) and how many nodes are under it. “Compliance settings for nodes” shows the org's switch, “Non-monitoring for the org” (On or Off): pressing it asks for a reason and your passkey (“Switch on with a passkey”, or off). “Per-node overrides” lists the nodes with their own settings; a row's “Change”, or “Change a node's settings”, opens a node's profile and its own switch the same way.

Under the profile Central Lite has only aggregates of a node's activity, and no cost per person. What still arrives: the hourly counts, status reports, and security and audit events with their time, the machine, the task id, process ids and the agent kind. On a single-user machine those still relate to one person, and there is no minimum count or coarser window yet: say so in your assessment.

Those events are what a task's incident timeline in Central Lite shows: one task's uploaded events, oldest first, with Central Lite's own actions on it, metadata only, never a prompt. Opening a timeline writes nothing, but every evidence bundle exported from it is chained with the admin who exported it and places a hold on the task's records. A kill order can name one person only by the pseudonym nodes upload (principal:kh1:), and cannot find a machine under the non-monitoring profile, whose uploads carry none; its reason is shown on every machine in scope, so it names no one.

If your identity provider provisions people into Central Lite with SCIM, Central Lite keeps only each person's userName, display name, externalId, whether they are active, their groups and, if you chose one, a login attribute; emails, phone numbers, addresses and every other attribute are dropped as the request is read. Who sees names in the console is a matter of role: viewer and finops never see people, and the other roles do. Put both in your assessment (what SCIM creates, what each role sees).

The console's Transparency screen: what nodes send, by data class, the retention in force and the compliance settings with the org's switch

See also

Investigate a task across the fleet

Read a task's timeline in Central Lite with its gaps, export a bundle the org root's evidence key signs, verify it offline and hold its records.

6. Publish the settings again after a restore

After Central Lite is restored from a backup and restore finish has run, its nodes can hold compliance settings newer than the backup's. Central Lite then does not sign the backup's settings on its own, since they could undo a switch made after the backup: it raises compliance_settings_behind_nodes and waits. Check compliance show, then compliance publish signs the settings as they stand, above every version a node reports, chained with your reason. On a Central Lite that is not behind, it signs the same settings again at a newer version.

garm-central compliance show
garm-central compliance publish --reason "settings checked after the restore"

Checked by the walk: this command ran as written on a Central Lite host.

See also

Back up and restore Central Lite

Write Central Lite's encrypted backups by hand and on a schedule, and restore one so that no enrolled node refuses what Central Lite signs next.