Agent governance

Agents you can let work, because every step can be traced

Organizations move from trying agents to running them. Policies and model registers exist. What is missing is the level below: which context an agent received, which suggestion it made, which source backs it and which person approved it. rygg records that for every statement.

Find your starting point Demo graph: Agent governance

Situation

Agents are in use in engineering, documentation and service. Nobody can say with certainty which knowledge they worked with.

Governance platforms register models and policies. They do not show whether a single answer rests on verified content.

Works councils, compliance and security ask the same questions about every new agent, and the answers are collected again each time.

Questions rygg answers

Which context package did this agent receive, and how current was it?
Which suggestions are still open, and who has to review them?
Which answers rest on statements that are outdated or contradicted?
Which roles are affected by this agent, and has that been agreed?
How often do agent suggestions pass review without changes?
Where did an agent decline to answer because evidence was missing?

Applications in this use case

Agent governance: 4 applications

context engineering with rygg

Context packages per agent

Which slice of the graph an agent may read, how current it is and how large it may be.

Agent context

Agent task: Every agent receives a defined package instead of the whole graph
Reads: Package definition: node types, status filter, size limit
Suggests: Request log with package version
Approved by: Platform owner

Metric: Answers given from the current package version

Review queue and decision rights

Agents write suggestions only. Who may approve what is defined per role and node type.

Agent context

Agent task: Places suggestions in the queue of the responsible role
Reads: Decision rights per role and node type
Suggests: Suggestion with sources
Approved by: Role with the right to decide

Metric: Time from suggestion to decision

Evals in operation

Traceability, abstention without evidence and respect for status, measured continuously, not once.

Agent context

Agent task: Eval agent runs test questions with expected answers against the graph
Reads: Test questions, expected answers, graph release
Suggests: Eval report per release
Approved by: Platform team

Metric: Pass rate per eval and release

View for works council and data protection

Which roles and tasks an agent affects, what has been agreed and what is still open.

Agent context

Agent task: Lists affected roles and tasks for a new agent
Reads: Roles, tasks, agreements, agent permissions
Suggests: Impact overview draft
Approved by: Works council and HR

Metric: Agents with an agreed impact overview

What rygg provides

Agent, permission, context package, suggestion and review as nodes in the graph, with provenance
Review queue: agents write suggestions only; people with the right to decide approve
Evals that run against the graph: traceability, abstention without evidence, respect for status
A view for works council and compliance: affected roles, tasks, agreements, open points

Who works with it

CTO, Head of Software Development
Platform and AI teams
Compliance and data protection
Works council

Metrics we agree on

Share of agent answers with complete provenance
Share of suggestions approved without changes
Correct abstentions where evidence is missing
Time from suggestion to verified result

What rygg is not here

Not a governance platform, model register or identity management. rygg connects to them and adds the evidence per statement.

How it starts

A pilot with one agent that reads a context package through an API, writes only suggestions and is measured with agreed evals.

Find your starting point

Usually starts withPilot 4 to 8 weeks

Four to eight weeks, one slice of your organization, one product line, one field or one agent task as a graph.

Demo graphAgent governance

Agents in engineering and service: context packages, suggestions, reviews, permissions and the view of the works council.