What is Stratos?

Stratos is an open-source platform for agents to author, use, and manage reusable workflows. An agent can capture a procedure in TypeScript, run it again, and compose it with other workflows instead of figuring out every step from scratch each time. Any agent can reuse a workflow, just like a skill.

An agent is useful when a task needs reasoning or exploration. But not every step needs a model to decide what happens next. Fetching known data, passing a result to another task, or choosing a branch from a validated answer can be ordinary code. Stratos lets you combine both approaches: code coordinates the work; models contribute where judgment is needed.

Small workflows become building blocks for larger ones. One agent might create a workflow that summarizes a document. Another can reuse it in a research workflow, and a third can include it in a weekly-report workflow without rewriting its implementation. The useful procedure lives in shared code, not only in the conversation that produced it.

What is a workflow?

A workflow is a program that takes input, performs a task, and returns a result. Its input and output have explicit schemas, so callers know what to provide and what to expect.

For example, an agent could create an issue-intake workflow that accepts a bug report, classifies the issue, gathers relevant context through a tool, and asks a coding agent to prepare an engineering brief when investigation is needed. The workflow’s code decides how those steps connect and what result to return.

Unlike a prompt describing a procedure, a workflow makes that procedure executable. You can inspect it, test it with representative inputs, and run the same coordination logic again. Model answers can still vary; putting the surrounding steps in code does not make a model infallible.

In Stratos, a workflow lives in a directory with a workflow.yaml file for its name and configuration, and a workflow.ts entrypoint for its schemas and run function. It does not have to use an agent at all: a workflow can be ordinary code, a single model decision, or a combination of many steps.

How do agents author and share workflows?

Agents are not only steps that workflows invoke: they can be the authors and callers of workflows too. Stratos provides CLI commands and an authoring skill so a coding agent can work through the whole lifecycle:

1. Discover and reuse. Find available workflows, inspect their input and output contracts, and run one that already fits the task.
2. Author and compose. When the task needs something new, write a workflow or combine existing ones into a larger automation.
3. Verify and maintain. Build the workflow, correct errors, check its results, and use run evidence to revise the saved code.

Workflows are files rather than private agent state. Agents with access to the same project’s .stratos/workflows/ directory can discover and reuse the same workflows. The user-wide ~/.stratos/workflows/ directory makes workflows available across projects to agents running as that user. Sharing a workflow shares its code and contract, not the conversation that authored it.

People can inspect, edit, and run workflows too, but do not need to hand-author each one. With the right context, any capable coding model should be able to use Stratos. The getting-started guide explains how to give an agent the bundled authoring guidance.

Why use Stratos?

Reuse a task without repeating its instructions

Once a task is a workflow, another agent or workflow can reuse its input and output contract rather than reproduce its internal steps. That makes it possible to assemble larger automations from focused pieces that can be understood and tested separately. TypeScript checks how those pieces fit together, and Stratos validates their inputs and outputs at runtime.

Keep predictable work in code

Use familiar functions, conditions, loops, and Promise.all to coordinate work. If you already know which tool to call, code can call it directly through a typed MCP client; MCP connects applications to external tools and data. Use an agent when the next steps require reasoning or exploration. You do not need to put the whole automation inside one long agent conversation or learn a separate graph language.

Use more than one kind of intelligence

Not every judgment needs a general-purpose LLM. Stratos’s Decision API integrates models such as Jev for focused decisions: choosing a category, assigning a score, or estimating whether a condition holds. A workflow can use one of these decisions to route a request, then invoke a coding agent only for the branch that needs deeper reasoning.

This can reduce cost and latency compared with using a full agent conversation for every decision. The benefit depends on the task and model; it is not a performance guarantee. Stratos also supports local models through Ollama and agents through Claude Code, Codex, and OpenCode. You choose the models and configuration appropriate to each step.

See what happened during a run

Stratos records local logs and agent events. Optional OpenTelemetry traces connect a parent workflow with its child workflows, model requests, and tool calls, helping you follow execution across the automation rather than inspect isolated conversations. The telemetry guide explains what is collected and how to view it.

How do workflows compose?

Calling a workflow from another workflow is called a child workflow invocation. The caller is the parent. The child keeps its own implementation and configuration; the parent passes input and receives its result.

Suppose an agent has created a workflow named summarize that accepts {text: string} and returns {summary: string}. Another agent can build a parent named prepare-brief that reuses it like this:

import createSummarize from '../.generated/summarize.js';

// Inside the parent workflow's run function:
const summarize = createSummarize();
const result = await summarize.execute({text: input.notes});

Here, input.notes is a string declared in the parent’s input schema. result.summary is the child’s typed output, ready for the parent’s next step. The parent does not need to know whether the summary came from an agent or another workflow.

The import points to a typed wrapper generated by Stratos, not to the child’s source code. Keep both workflows in the same discovery root, for example under .stratos/workflows/, and build the parent:

stratos build prepare-brief

Stratos prepares the child and generates the wrapper automatically. The complete child-workflow example shows the surrounding parent definition using the greeting workflow from the authoring guide.

Each execute call starts a fresh child run with its own validation and resource cleanup. Its result goes back to the parent; errors reject the call and can be handled with ordinary try/catch. The same child can also run independently from the CLI. Reuse it in a loop, call independent children in parallel, or pass one child’s result to another to build up a larger task.

Start building

Start with Getting started to install Stratos on macOS or Linux with Node.js 24 or newer and make its authoring guidance available to your coding agent. Give the agent a task and ask it to find or create a workflow for it. Author a workflow walks through the files and build-and-run cycle using a small example that runs without model credentials or external tools. The primitives guide explains how workflows add agents, tools, and structured decisions.

Stratos is early-stage software. Workflow code runs with your process’s permissions: it is not a sandbox. Agents and tools can change files, call external services, and incur charges. Review code, credentials, and permissions before running a shared workflow, and treat logs and traces as potentially sensitive.

See current runtime limitations before adopting it for a task. These docs follow the repository’s main branch, which may include changes newer than the installed npm release.