What Little Workflow Is

Little Workflow is a TypeScript library for AI-generated work that must be bounded, durable, inspectable, and replayable.

The developer describes a goal, inputs, schemas, model slots, tools, skills, and safety rules. An AI agent running in a structured Harness designs the workflow graph (Planner) and executes the resulting steps (Worker). A Fixer can repair a failed code.run step, but only with a custom worker harness: the default worker does not run code.run steps. Every run is observable through a unified event log and durable through per-step memoization, and filesystem Memory Stores accumulate across runs. To compose several workflows, hand them to a Little Harness agent as typed tools.

Little Workflow is optional. Most projects start with Little Harness and add a workflow when a job must be bounded, durable and replayable; see How they fit. To create a project with an agent and an optional sample workflow, use Setup.

pnpm add @little-stack/workflow@alpha ai@^7 zod @ai-sdk/anthropic   # or any AI SDK 7 provider

The current release is @little-stack/workflow 0.2.0-alpha.2, on the alpha and latest npm tags. It requires Node.js 22+ and AI SDK 7, which is a peer dependency. pnpm 10 projects must allow the better-sqlite3 build; see the Quickstart.

WorkflowDefinition
-> Harness session (Planner)
-> planner draft LWIR
-> deterministic validation and repair
-> immutable WorkflowVersion
-> durable Run (workers execute each step via Harness)
-> artifacts, events, streams, replay, retry, fork

Key Primitives in v0.1.0-alpha

Harness is the agent runtime each role lives in — model + bash environment + scoped storage + tools + persistent agent loop. Think Claude Code or OpenCode. The SDK ships a default workflowHarness; devs can substitute by implementing the single run(task, ctx) interface. Every model call and tool call inside a harness session is event-logged for deterministic crash recovery.

Composition happens on the Little Harness side. asHarnessWorkflow turns each workflow into a typed tool for a createHarness agent, which decides when to call which workflow and how to chain their outputs. There is no static pipeline IR. See Compose With Little Harness. The older Orchestrator role (runWorkflow with an array of definitions and the plan_workflow / run_workflow / start_workflow tools) is deprecated.

Memory Stores are persistent, filesystem-shaped collections of text documents mounted into each harness session at /mnt/memory/workflow/, /mnt/memory/pipeline/, and /mnt/memory/org/. Planners write lessons learned; future planners read them. Agents self-improve across runs without model fine-tuning.

Product Boundaries

Little Workflow is:

an AI workflow compiler
a durable execution runtime with harness-level crash recovery
an artifact-first audit system
a capability allowlist over model, tool, skill, and code access (the full firewall is post-alpha)
a local development experience through Local World

Little Workflow is not:

a chat UI
a generic multi-agent framework
an n8n clone
a Temporal replacement
a low-code automation builder first

Where It Sits In The Stack

Little Workflow is the SDK half of the Little stack, with Little Harness. A workflow's planner, worker and fixer roles run in a harness, so a workflow run lives in the same workspace as the agent that started it. LittleDB, the product, records each run as a trace, scores it, and improves the workflow's step prompts and models the same way it improves an agent; db.tracingWorld() and db.managedWorkflow() are the two hooks. Workflow needs nothing from LittleDB or from a workspace engine to run. See How they fit.

Start Here

Setup — create a project with little setup (shared guide for all products)
How they fit — Workflow next to Harness, LittleDB and workspaces
Releases — what 0.2.0-alpha.2 contains and what remains unpublished

The Guides and API section keeps the v0.1.0-alpha segment in its URLs so old links stay stable. That segment is a historical URL, not the package version. The section covers the current alpha line, including the Harness, Memory Stores, Run Scratch, and Skills primitives, plus composition with Little Harness. See the Roadmap for post-alpha direction, and Comparisons for where Little Workflow fits next to other tools.

For authoring, use defineWorkflow: one model, inline tools, and plain input/output schemas, with the full config (several model slots, a planner config, a fixed plan) available on the same call. Start with the Quickstart.

Overview — Harness-based agent runtime, Memory, Skills
Quickstart — define a workflow and run it in one call
Define a Workflow — the ergonomic defineWorkflow front door
Compose With Little Harness — hand workflows to an agent as typed tools
Roadmap — where Little Workflow is headed after the alpha