What Archi is
Archi automates operations for scientific computing: experiment computing operations, analysis user support, data quality and cluster management. Agents watch the system and propose actions. Once a person approves a proposal, a separate component carries it out; in 3.0 no automation acts on its own. Every step is grounded in your facility's own knowledge, and its result is written back into that knowledge.
Archi 3.0 ships the context layer, the automation runtime and two
automations. ticket-triage triages new and
updated Redmine, JIRA and ServiceNow tickets: it drafts a cited reply for a
person to send, or proposes a status change, assignment or reply that runs
only after a person approves it.
instance-health watches the Archi
instance itself.
Archi is built on OKG, a knowledge graph with a publish step, a set of read tools and a review ledger. You do not need to know OKG to start. This site explains each OKG idea where Archi uses it.
Note
This is the 3.0 documentation. The earlier chatbot is documented under v2 (legacy). See Moving from v2 to 3.0.
Two layers
Automation is only as good as what the agent knows. So Archi has two layers.
- The context layer holds what your facility knows: documentation, tickets, catalogs, monitoring, code and the record of past operations. It has to be solid. That means current, complete for the questions automations depend on, linked across sources, cited on every read, and measured.
- The operations layer runs automations over that context. Each one follows the same loop: signal, context, proposal, decision, action, outcome. Actions run under explicit autonomy levels, up to L2 in 3.0. Outcomes go back into the context layer, so the next triage can see what happened last time.
The working rule is: algorithmic first, agents for judgment. Fixed checks (invariants, freshness, thresholds, evals) decide everything they can. Agents handle what needs reading and judgment.
What you install
You install a bundle: a named selection of packs. A pack is the smallest installable piece, such as one source family or one automation. Four bundles ship today:
| Bundle | For |
|---|---|
cern-team |
Any CERN team: documentation and meetings, with no credentials needed to start. |
cms-physics |
CMS physics: analysis notes, published summaries, physics TWiki pages and the analyses that link them. |
cms-comp-ops |
CMS computing operations: sites, storage, datasets, workflows, site status and past incidents. |
cluster-manager |
Any cluster operator, at any site: documentation, tickets, operational memory, and optionally Slurm, HTCondor, ServiceNow and node health. |
The library lists every bundle and pack with what it
reads. Your own sources and automations can ship as packs from your own
package. Start with the quickstart: it takes an empty machine
to a cited answer with the cern-team bundle.
Parts you can replace
Most parts can be replaced, and the defaults work out of the box. The knowledge graph and the automation runtime cannot be replaced in 3.0.
| Part | Default in 3.0 | How you replace it |
|---|---|---|
| Knowledge graph | OKG (AKMON) | Not replaceable in 3.0. The MCP tool surface (search, expand, inspect, filter, map, aggregate) is the contract another store would have to meet. |
| Ingestion | Archi connectors | Any pip package that registers packs through the archi.packs entry point. See Contributing a pack. |
| Tools | The instance's OKG MCP server plus Archi live tools | Any MCP server declared in the instance. See Connecting an agent. |
| Automation runtime | OKG deployment workflows | An external runner using the proposal API. This waits on an upstream OKG change. |
| Agent and chat | Open WebUI, built in | Anything that speaks MCP: Claude Code, Codex, OpenClaw, your own. |
| Language model | Chat: a local ollama model (default llama3.1:8b). Agents over MCP: the agent's own model. |
Chat: --answer chat_model=, --ollama-url, or a connection added in Open WebUI. Agents: their own configuration. Archi stores no model API key. |
Who it is for
First, operations teams: CMS computing operations, CRAB support, WisDQM and cluster managers at any site. Second, a team that only wants the context layer and an agent over it in an afternoon. That team installs the same bundles with no automation selected.
The design is in ADR 0002.