How they fit

The Little stack is three things that fit together.

• LittleDB is the product. It records what your agents and workflows do as traces, scores them (code checks, AI judges, people, and what your users said), and improves them: a reviewer reads the traces and proposes a change, LittleDB tests the change on your own past conversations, checks it against your guardrails, and puts it live or asks a person. LittleDB is private source for now, and Harness and Workflow run fully without it.
• Little Harness and Little Workflow are the SDK. You build agents with Harness, and bounded, durable jobs with Workflow. LittleDB traces both and improves both. Most projects start and stay with Harness; Workflow is optional.
• Workspaces are where an agent's work lives: each user's own folder, kept between visits and ready when they return. Managed workspaces come with LittleDB's hosted service; the self-hosted path is Little Compute. See Workspaces below.

your app or the CLI
  -> Little Harness (agent runtime)
       -> workflows option -> Little Workflow (optional; depends on Harness)
       -> workspace: folder tier in process, a container when a command needs one
            managed: LittleDB's hosted service   self-hosted: Little Compute
       -> onEvent, feedback, eval worker -> LittleDB
       <- resolve (which version to run)  <- LittleDB

The same agent folder (agent.ts, with tools/, skills/ and workflows beside it) runs unchanged on a laptop or a Podman box today, and on the managed tier as it lands; only the sandbox builder differs. That portability, the merge-back of parallel children's work into one workspace, and the loop from real user outcomes back into the agent's configuration are what the stack is for. The runtime underneath is deliberately thin.

Start with Harness

A Harness agent needs a host, a model, instructions and tools. Little Harness creates a private workspace per session, stores large outputs as artifacts, records a trace, and returns a result. See What Little Harness Is.

An agent can live in an agent folder (agent.ts, with tools/ and skills/ beside it). The little-harness CLI loads that folder, and so can your app through loadHarness(). Projects created by little setup use the same agent definitions for the CLI and for the generated web route.

Add Workflow for bounded, durable jobs

Use Little Workflow when a job must stay inside declared limits, survive a crash, and be replayed or inspected step by step. An agent designs the workflow; Little Workflow validates it into an immutable version and runs it durably. The @little-stack/workflow package depends on @little-stack/harness, and its planner, worker and fixer roles run in a harness.

To let a Harness agent decide when to run a workflow, adapt it with asHarnessWorkflow and pass it to createHarness({ workflows }). The canonical guide is Compose With Little Harness.

Send it to LittleDB

LittleDB is one service, reached with one URL and one API key. Its client resolves which version of an agent or workflow to run, sends Harness turns and Workflow runs as traces, forwards what your users said, puts agent-made workflows behind review, and runs LittleDB's experiments in your app. That is how the loop closes: real outcomes come back as a better version of the agent. See LittleDB.

Workspaces

A workspace is one user's folder and the shell over it. Most turns run on the folder tier: Harness Tier-0 just-bash over the user's files, in process, in milliseconds, with the command classifier deciding before anything runs. A command that needs a real machine (git, a package manager, a build, tests, a long-running process) escalates to a container that is attached to the same workspace and sleeps again after ten idle minutes. The workspace, not the container, is the source of truth; the container is borrowed.

Two ways to run one:

• Managed with LittleDB's hosted service (private for now): LittleDB holds each user's workspace, the Harness folder tier runs for most turns, and a container wakes only when a command needs it. Each turn reports its tier and infrastructure cost to LittleDB.
• Self-hosted with Little Compute on rootless Podman: the same workspace with durable Tasks, fork, and merge-back of parallel children's work. Little Compute's control plane is frozen as of 2026-10-23; what is merged keeps working, and nothing new is added to it.

In Harness terms both are a sandbox: the folder tier is justBashSandbox(), and the container is the nextTier of tieredSandbox(). On a laptop the same agent runs with Tier-0 alone or a local Podman as the next tier. Little Workflow does not depend on either workspace engine.

Runs on your machine

• Harness sessions, traces and artifacts are stored by the Local Host in a folder you choose.
• Workflow runs are recorded by Local World in an embedded SQLite store on disk.
• A self-hosted workspace runs on rootless Podman; see Self-hosting. It is not publicly installable yet.
• LittleDB is optional. Its dev mode must not be exposed to untrusted networks; see LittleDB.

Sessions, runs and tasks

A Harness Session is one agent conversation or job. A Workflow Run is one durable execution of a workflow version. A Compute Task is one unit of agent work that a self-hosted workspace schedules, with retry Attempts that use Sandbox Leases for isolated execution. See the Glossary.