Digdir's agent platform: agentctl runs long-running agents in isolated sandboxes, built on a reusable Sandbox SDK.
Agents get their own computer, with mediated network access and no real secrets inside the sandbox.
This README is a short placeholder while the repository is being cleaned up. A full rewrite with an architecture diagram follows.
| Path | What it is |
|---|---|
agentctl/ |
The agentctl CLI and the agentd session API backend, independent of model and harness. Start here to install and run an agent. |
sandbox/ |
The Sandbox SDK, with an implementation on Digdir's forks of microsandbox and libkrunfw. |
images/ |
Container images for building agent images. |
agents/self-dev/ |
The agent that develops this platform. It is also the example to start from when you make a new agent. |
integrations/ |
Slack and GitHub listener from the earlier orchestrator pipeline. It is kept as a possible entry point to agentctl agents, but nothing consumes its events right now. |
docs/presentations/ |
Slide decks on agentctl and the Sandbox SDK, including a security view. |
Install agentctl and run an agent as described in agentctl/README.md. To work on the
platform itself, use the self-dev agent: see agents/self-dev/README.md.
Until October 2026 this repository also held an orchestrator pipeline: a set of agents fed through a file contract
from integrations/, an LLM gateway, host scripts and a knowledge base under doc/. It is archived as the tag
v2.0-pre-cleanup.