Announcing MagicStride

Most agent work dies with the runtime that produced it.

A conversation runs, a process exits, and the context — the plan, the half-finished change, the reasoning, the artifacts — goes with it. The next session starts over. Teams rebuild context instead of building on it. We have been treating agents like chat windows when they need a workspace.

We are building MagicStride to change that.

MagicStride is a persistent, sandboxed, team-oriented platform for working with AI agents across technical and business workflows. The work survives the runtime. The runtime does not own the work.

What MagicStride gives you

Persistent workspaces Projects, sessions, checkpoints, and artifacts that outlive any agent process. Replace the runtime, not your progress.

Coordinated execution Decompose an outcome into focused tasks, run them in parallel, verify the outputs, and iterate toward completion.

Autonomy with boundaries Each task gets an isolated sandbox and explicit access limits. Human-in-the-loop and autonomous workflows share one task model.

A durable record Canonical events, audit history, and reviewable artifacts stay in the platform domain — not in a process that can be killed.

Team governance Work belongs to organizations, teams, and workspaces — not to personal process memory. People stay at the consequential boundaries.

How it works

1. Import a Git repository as a resource and create a persistent workspace.
2. A Runner spins up an isolated sandbox, materializes a checkout, and starts the agent.
3. Stream messages, tools, approvals, files, and diffs through the harness.
4. Inspect outputs, checkpoint project state, and dispose of the sandbox.
5. Recreate the environment later and resume the session — commit, push, merge, and deploy remain separately authorized actions.

Who it is for

Engineering, research, operations, finance, policy — anywhere a team needs continuity, accountability, and a reviewable record from agentic work, not just a transcript.

Where we are today

MagicStride is currently in pre-alpha. The conceptual architecture — what we call Foundry — is established: the Clients → Harness → Runner → Sandbox → Server path, the persistent-versus-disposable state model, and the domain vocabulary are all decided. Solution architecture is in progress.

What exists today is a public website, the architecture documentation, and a fully scripted agent-client playground with two scenarios: a human-reviewed financial close and an autonomous music-video production run. No authentication, model execution, sandbox provisioning, or external integrations are connected yet.

The recommended first runtime milestone is narrow on purpose: one local Git project, one local Runner, one sandbox provider, disposable instances. We are validating the lifecycle — start, stream, approve, stop, resume — before widening the surface.

If you are doing high-stakes work with agents and want a home for it that lasts longer than the process, we would love to hear from you.

Explore the platform → · Read the architecture → · Try the console demo →