Skip to main content

Concepts Overview

Cordum is an agent control plane — a governance layer that sits between autonomous agents and the real systems they act on. This page is a map of the load-bearing concepts. Each card links to the page where the concept is defined in depth.

The four things Cordum governs​

  1. What an agent is allowed to request → Safety Kernel · evaluates every job pre-dispatch, returns ALLOW / DENY / ALLOW_WITH_CONSTRAINTS / REQUIRE_APPROVAL.
  2. What an agent is allowed to say (model output) → Output Safety · post-execution scanner pipeline that allows, redacts, or quarantines results.
  3. What tools an agent is allowed to call (MCP) → Agent Protocol (CAP) and the MCP governance surface.
  4. Who gets to approve the edge cases → approvals queue, surfaced in the dashboard and cordumctl.

Architecture anchors​

  • System Overview — the service graph (gateway, scheduler, safety kernel, workflow engine, context engine, dashboard).
  • Agent Protocol (CAP) — the wire contract. CAP = "MCP for agent behavior" (MCP = say, CAP = do). Spec lives in the cap repo.
  • Packs — reusable policy + workflow bundles that target a domain (finance, ops, data).
  • Context Engine — context windows, memory pointers, Redis-backed artifact storage.
  • Scheduler Pool Spec — how jobs route from the bus into worker pools.

Safety & governance deep-dives​

Operator-facing reference​

Everything related to running Cordum in production lives under Operations — install, helm, deployment, config reference, horizontal scaling, performance tuning, and enterprise features.

Developer-facing reference​

Glossary​

For a one-line definition of every Cordum term (Safety Kernel, Pack, CAP, BusPacket, Context Engine, Scheduler, Output Policy, Remediation, Saga, Trust Gap, etc.) see the Glossary.