Connect your agents and tools.
Add adapters that call your existing agent runtime and tools. Connect Jev for typed decisions. Keep credentials and permissions in your host.
BACKED BY Y COMBINATOR
Add Jev-powered
workflows to
your agents.
Less intelligence wasted.
More accurate results.
Let your agent solve the job. Turn the repeatable steps into a workflow. Use Jev to route decisions and your existing agents to do the work.
BUG REPORT 001“Checkout crashes for every customer.”A new bug report
BEFORE / AGENT LOOP
Every bug. Another triage loop.
Your agent works out severity, ownership, and how to file each issue.
Simulated demo
01 / LESS INTELLIGENCE WASTED. BETTER RESULTS.
Stop paying to solve
the same problem again.
Use the agent where it adds value. Let Jev handle focused decisions and code handle the rules. Research only what can change the outcome.
REPORTED AML CASE STUDY
Full agent
$2.89per runAgent + workflow
$0.25per runThe same procedure, with focused Jev decisions and agents working where they add value.
90%
lower model costReported reduction, rounded
MODEL COST / PER RUN
1,000 runs · arithmetic illustration
$2,890$250
ACCURACY IN THE TUNING RUN
90%95%
First 100 alerts → final 300 alerts.
Different stages of the 1,000-alert tuning run.
Reported AML case study: cost comparison on 100 fresh alerts; accuracy from the separate tuning run. Volume figures multiply the reported rates and exclude learning costs and operating overhead. These are not pricing or guaranteed results. Read the experiments ↗
02 / FROM YOUR AGENT TO A WORKFLOW
Keep your agent.
Give it a repeatable process.
Your harness keeps its loop, tools, permissions, and budgets. AgentRun defines which steps run, which can run in parallel, and when to stop or hand off.
Define how the work flows.
Describe the steps in TypeScript or JSON. Route a request to the right agent, check its output, and make the next step explicit.
Test it. Make it callable.
Test individual decisions and the whole workflow. Give your agent the workflow as a tool, or call it directly from your application.
// research-one-question · node excerpt// Full schemas and workflow in the example.{ node: 'sift', label: 'screen-evidence', itemsPath: 'search.sources', state: { subquestion: '{question}' }, out: 'EvidenceDecision', as: 'evidence', keep: { path: 'answersQuestion', gte: 0.8 }}Your agents can run inside a workflow.
Your agent can also call a workflow as a tool.
03 / THE BUILDING BLOCKS
A few clear steps.
A whole procedure.
AgentRun’s workflow language—the DSL—gives each kind of work its own place. Every step has a job. Every handoff has a contract.
Agents
Gather evidence, investigate a gap, or write a report. Each agent runs with the model, tools, and budget your harness provides.
Jev decisions
Route, classify, sift, or pick over structured data. Keep probabilities alongside answers so uncertainty can guide what happens next.
Ordinary code
Parse, compare, count, and apply rules. Call a tool, poll within a deadline, or check an artifact. Save model turns for the work that needs them.
You define the control flow.
Chain steps. Map over a list. Reduce the results. Loop within a bound. Escalate when the ordinary path cannot settle the case.
chain→map / reduce→loop until→escalate
04 / IMPROVE THE PART THAT MATTERS
See the decision.
Change one piece.
An explicit workflow gives you smaller parts to evaluate, improve, and reuse.
Follow the reasoning path.
Read inputs, typed answers, probabilities, and routing rules at each step. See which part of the procedure led to the outcome.
Test a decision on its own.
Freeze a node’s inputs and test a new model or rubric. Replay deterministic rules over saved answers without repeating research.
Build once. Compose again.
Give a workflow its own input, output, and tests. Reuse it as a component inside another workflow or a tool inside your agent.
05 / WORK WITH PI
Your coding agent.
A workflow vocabulary.
The AgentRun adapter runs typed workflow nodes with pi. It can also help author workflow candidates with schema validation and independent acceptance checks. The core workflow runtime can run without pi.
The native pi extension is a design preview: build, test, and inspect workflows without leaving pi. Try the working SDK and CLI from source today.
❯
Turn this procedure into a reusable workflow. Use agents for research, Jev for decisions, and code for the rules.
A WORKFLOW YOU CAN REASON ABOUT
Explicit steps and routing
Structured inputs and outputs
Small pieces you can evaluate
Native extension planned · SDK and CLI available from source
06 / BUILD WITH AGENTRUN
Read the code.
Run the real thing.
Full documentation and working examples from the repository, here in one place. Start with scripted fixtures, then connect your own tools and models.
01 / DOCUMENTATION
Quickstart
Run the interpreter with fictional sources and scripted responses. No model keys needed.
02 / DOCUMENTATIONAuthor a workflow
Define typed contracts, reuse a component, and test a decision independently.
03 / DOCUMENTATIONInteractive examples
Inspect recorded calls and follow the actual execution path.
04 / DOCUMENTATIONWorkflow reference
Explore primitives, schemas, events, and execution limits.
05 / DOCUMENTATIONConnect your agent
Bring your tools, model access, permissions, and budgets.
06 / DOCUMENTATIONJev + pi adapters
Connect real typed decisions and agent loops.
Node 22.19+ · Release candidate · npm packages not yet published. Release status →
07 / SEE IT IN PRACTICE
From an agent run
to a reusable workflow.
Watch the AgentRun introduction for the full idea: learning a job, expressing the procedure, and optimizing repeatable work.
agent.run()INTRODUCING AGENTRUNLearn the job.
Write the program.
Write the program.
Introducing AgentRun · GREP AIWatch on YouTube
START WITH THE AGENT YOU HAVE
Give your agent
a better way to repeat.
We’re preparing AgentRun for open source.
Explore the source, run a workflow, and help shape what comes next.