AGENT WORKFLOW LANGUAGE

Prompts make wishes.
Code makes them real.

Define steps, outputs, and acceptance criteria in a .flow file. Agents do the work. The runtime validates results, limits resources, and stops when conditions aren’t met.

Build your first workflow See the language

Standalone language · Embeddable runtime · Model independent

verify.flowA code delivery workflow

Implement the changeagent(implementer)

stage

02a

Run the testsexpect TestReport

02b

Review the risksexpect ReviewReport

Evaluate acceptanceTests pass. No blockers remain.

verify

Acceptance ruleRUNTIME

pass when tests.passed
 and review.blockers.empty

Conditions not met? The workflow stops here.

Workflow illustration · from verify.flow

Before executionCheck syntax, types & dependencies

After each callValidate the agent’s output structure

During executionLimit concurrency, access & time

01 / MAKE THE RULES EXPLICIT

“Follow these steps”
only gets you so far.

A single model call rarely needs a new language. But once tasks have dependencies, run in parallel, and need acceptance checks, the workflow deserves an explicit definition.

Explore the design decisions

Outputs have a contract.

Declare the fields and types your next step needs with expect. Results must pass structural validation before they can be used as that type.

expect Result

Progress has a condition.

Write acceptance criteria as a pass when expression. Use require to stop failed results from moving forward.

require result.passed

Execution has boundaries.

Declare concurrency, call budgets, timeouts, and requested permissions. The host policy still determines which capabilities are available.

limits · tools · timeout

02 / READ THE CODE

The rules live outside the prompt.

Three excerpts. Three places where the runtime enforces your intent.

simple.flow / Code excerpt

1stage answer -> Result {
2  return agent(writer) {
3    task "Answer the request"
4    input {
5      request: input
6    }
7    tools none
8    expect Result
9  }
10}

expect Result generates a JSON Schema for runtime validation. One repair attempt is allowed by default; invalid output then raises an error.

Read the language guide

Surrounding types and some task parameters are omitted for clarity. See the docs for complete examples.

03 / START WITH YOUR USE CASE

Complex work. Clear handoffs.

Browse all examples

Code delivery

Implement → Test & review → Verify

Separate implementation from verification. Run checks in parallel and let the acceptance rule decide what proceeds.

verify.flow

Batch processing

Task list → Bounded fan-out → Collect

Dispatch each item, limit how many run at once, and collect structured results for the next step.

parallel.flow

Plan and execute

Plan → Assign → Deliver

Constrain a plan with types, item counts, and uniqueness checks so the next stage gets explicit work items.

delivery.flow

START SMALL

Run one step.
Then build the workflow.

Start with simple.flow. Check the syntax, inspect the compiled output, and run it in the sandbox. Connect a real model through your own AgentRuntime adapter.

Follow the quickstart

Run from the repository root

$ pnpm install
$ pnpm build
$ node packages/cli/dist/main.js check examples/simple.flow

Requires Node.js 20 or 22 and pnpm ≥ 9. The CLI demo uses FakeAgentRuntime; no real model is called.

For workflows with stages, dependencies, and acceptance criteria. For a single question, a direct model API call is enough.

Prompts make wishes.Code makes them real.

“Follow these steps”only gets you so far.