Agent-native business platform

Plan the business.
Agents do the work.

Siteplan is a CMS and application foundation built on Django and Wagtail. It treats an AI agent as a first-class actor next to a human — sharing the same event stream, the same workspace and the same audit trail, while holding its own scoped authority.

One Django ASGI process · Postgres as truth · Redis for fan-out · HTMX and Alpine on the front end

Workspace · three columns
COLUMN 1 Global shell PINNED ARTIFACTS COLUMN 2 Control and stream APPROVAL REQUESTED RUN · DONE · 12.4K TOKENS COLUMN 3 Canvas GEN_UI BROWSER DIFF LOGS VOLUME BY REGION ARTIFACT v2
Routes · stream · canvas Server-rendered fragments
The problem

A chat box bolted onto a form is not an agent-native product.

Most "AI features" in business software are a sidebar: the agent cannot see the work, cannot act on it, and leaves no accountable trace. Siteplan starts from the opposite premise — the agent is an actor in the system, with identity, permissions, limits and an audit record.

Separate the stream from the state

The event log records what happened. Canvases, artifacts, tasks and pins are rebuildable projections of it — never the other way round.

Give agents capabilities, not logins

Agents never authenticate, never receive staff flags and never hold a socket. They are invoked, checked at append time, and granted coarse, expiring capabilities.

Make the output durable

A conversation ends. An artifact does not. Publishing promotes a canvas version into a pinned, refreshable, viewer-scoped report that outlives the topic that produced it.

The workspace

Three columns, one continuous piece of work.

The shell never becomes a wall of chat. Column one is where things live, column two is how work happens, and column three is what the work produced.

  • Global shell — routes, rooms, pinned artifacts and reports.
  • Control and stream — the topic log, run activity and approvals.
  • Canvas — generative UI, browser, diff and log views of the current state.
Column 1
Global shell
routesrooms pinned artifactssaved views
Column 2
Control and stream
topic treerun log approvalscomposer
Column 3
Canvas
GEN_UIBROWSER DIFFLOGS
One shell, three lifetimes. A canvas is scratch state, an artifact is durable content, and a pin is a placement. Conflating them is how chat products lose the work.
The central loop

From a missing report to a first-class artifact.

This is the acceptance scenario the whole design is written against. If a step needs a concept the domain model does not have, the model is wrong — not the scenario.

01
Ask for work, not a chat
A task is created in a room, assigned to an agent user, and linked to a measurable objective. The assignee is a User, so assigning work to an agent is the same operation as assigning it to a colleague.
02
Query data without SQL
The agent requests a dataset by id. The server checks its capability, intersects the agent's data scope with the requester's, and executes a validated spec — never a query string.
03
Author on the canvas
The agent emits a generative UI mutation whose widget binds to that dataset. The server renders the fragment, appends the event, and the canvas becomes a version.
04
Publish a version
Publishing promotes the canvas version into an ArtifactVersion and appends ARTIFACT_PUBLISHED. The canvas is not consumed, so refinement continues.
05
Pin a placement
A pin references the artifact, not a version. Pin the same artifact to a team and to the whole organization and both placements follow the newest version.
06
Render per viewer
The report refreshes on a schedule. Every viewer's rows are resolved as that viewer, so two people opening the same pin can legitimately see different data — and a pin never grants access.
Position

The platform is the plan, not the prompt.

Siteplan is installed on-premise and built as a foundation, not a hosted toy. The architecture is chosen so that the same application can run locally under a trusting administrator, in a sandbox, in a container, or on a remote VM — by policy, not by rewrite.

Agents are powerful because they are bounded. Every irreversible action is approval-gated, every run is attributed to a human owner, and every tool call passes through one mediated dispatch.

Read the vision
Trust is layered
ORGANIZATION POLICY AGENT GRANTS · EXPIRING CAPABILITIES PLACEMENT RUN · WORKDIR + HOME ROOT DATASET · QUERY AS VIEWER MEDIATED TOOL DISPATCH APPROVAL REQUIRED RUN-SCOPED TOKEN DEFAULT DENY
Capability · data scopePlacement · isolation

An organization, its teams, its rooms — and agents that earn their place.