Parcha-ai/agentrun

What it solves

AgentRun tackles the problem of building agent workflows that are reliable, inspectable, and recoverable. Without a structured harness, agent steps are often ad hoc and hard to audit, especially when they call external tools or make decisions. AgentRun lets you define steps as a workflow document, gate progress with Jev typed decisions, and run the workflow durably so you can inspect, rerun, and recover it.

How it works

You write workflows as JSON or use the TypeScript builder with Zod contracts. Each workflow defines schemas for inputs and outputs, steps like call, judge, and agent, plus thresholds and review paths. The interpreter executes steps, and you supply adapters: runEffect for tools, runNode for your agent, and createJevRunner as runJudge for decisions. The workflow can call other workflows, map work in parallel, and run bounded loops. AgentRun also ships packages that extend the harness: pi-vision for reading documents, images, and video; pi-browser for giving the agent its own browser; pi-transport for running steps on different machines; and more. Everything runs on pi-durable for durability, and demos are available with scripted responses and no API key.

Who it's for

Developers building agents on Pi (versions 1.0.4 to below 1.4) who need structured, inspectable, and recoverable workflows. It’s also for teams that want to plug in their own tools, models, and decision logic while getting features like vision, browser control, and machine-transport execution out of the box.

Highlights

• Workflow DSL with typed schemas and runtime validation of state paths
• Jev typed decisions with confidence thresholds and review escalation paths
• Durable execution via pi-durable (commits each turn and tool call to storage)
• Pi extension with /agentrun commands for building, inspecting, and running workflows
• pi-vision: answer questions about documents, images, and video with evidence
• pi-browser: agent gets its own browser; you can take over and hand it back; pages filed as evidence
• pi-transport: run steps on local machine, Docker, Daytona, or Modal behind one Box interface
• pi-agent-learn: an agent that learns in its own git repository, curating notes into commits