Facility’s next direction: persistent workspaces and agents as code

[adrian-lorenzo]

We are changing how Facility represents and runs agent work.

Today, agent work is organized around short-lived runs, separate execution contexts, and several governance layers. We are moving to a simpler model: one persistent workspace for each story.

That workspace will retain the conversation, Git worktree, uncommitted changes, development services, and native Claude Code or Codex sessions until an authorized user explicitly deletes it. Compute may sleep or be replaced without losing the working state.

Agents will be configured in the repository under .agents/. The initial crew will include Architect, Builder, PR Reviewer, Address Review, CI Doctor, and Security Audit. Every agent will declare its engine, model, prompt, and triggers through the same file format.

Scheduled and GitHub-triggered agents will remain part of Facility. They will use the same persistent stories, workspaces, and configuration as interactive agents.

We are also simplifying access. Every configured agent will receive full access to its workspace and maintainer-level access to the repositories connected to the project. Authentication, project membership, tenant isolation, branch protection, and CI remain the relevant boundaries.

MCP will become the primary interface for automation and coding agents. The Facility UI will remain available for configuration, conversations, previews, status, and workspace management.

The proposal has been accepted and implementation is already in progress under #272. The work is assigned to @adrian-lorenzo, who is leading its execution.

The epic and its subissues are the source of truth for scope, technical decisions, dependencies, and progress. This announcement is published to give contributors visibility into the direction Facility is taking.