agentops / ecosystem
bounded automation, explicit authority

A system of owned state, not one omniscient agent

Work that survives the session.

The AgentOps ecosystem turns intent into scoped execution, then preserves the state, evidence, knowledge, and recovery path required for the next operator or agent to continue safely.

Who owns what?

Select a domain to inspect its authority, artifacts, and boundaries. Repository boundaries follow state ownership; shared visibility does not imply shared write access.

Trace one change

Switch between a human-led session and queued dispatch. The route changes; the proof obligations do not.

Authority matrix

The useful question is not “which tool can see this?” but “which contract may change it?” The table is intentionally asymmetric.

SystemOwnsMay mutatePortable projectionMust not become
sprintctlwork, claims, handoffssprint stateresume / handoff / renderproject manager or swarm lock
kctlknowledge review lifecycleits candidates and entriesknowledge Markdown / JSONbacklog writer
actionqaction and session lifecyclequeue state through actionctlappend-only eventsdirect repository editor
actionq-dispatchone bounded execution cycleisolated worktree + contract writesbranch, result, gate evidenceunbounded autonomous loop
auditctlrepo-local audit evidenceaudit index and shardsdaily NDJSONhidden central ledger
cockpitcomposed operator viewonly via owner APIsread models and linkscockpit-shaped monolith

Try the cockpit view

This public demo uses synthetic state to show how the operator surface composes repository, sprint, claim, dispatch, and audit projections. It has no connection to the private deployment and intentionally exposes no write controls.

agentops / cockpitsynthetic read-only demo

Active work and claims

Dispatch and audit feed

Boundary: this screen renders projections. A real mutation must return through sprintctl or actionctl with the required ownership proof.

Why the parts exist

Every component exists to remove a concrete ambiguity. Otherwise it is distributed-systems cosplay.

01 / ContinuityContext is generated from live state.

Resume bundles, claims, refs, and handoffs replace “remember what the previous agent meant.”

02 / ContainmentExecution is narrow and inspectable.

One action, one worktree, explicit path ACLs, pre/post gates, and a branch that survives failure.

03 / LearningHistory is not automatically knowledge.

Events remain evidence. kctl requires review before durable or coordination lessons become published artifacts.