AgentRun 0.1.0-beta.4 release notes / what's new

AgentRun 0.1.0-beta.4 is a workflow language designed to add structure to existing AI agents. It allows developers to define repeatable steps, integrate focused decision-making via Jev, and trigger agent investigation when specific criteria are not met. The system is designed to maintain the agent's existing tools, permissions, and model access while providing a formal framework for execution.

Core Architecture and DSL

The AgentRun DSL enables the definition of workflows as JSON or via a TypeScript builder with Zod contracts. This approach ensures that input and output types are inferred and intermediate state paths are checked at runtime.

Key architectural features include:

  • Node Types: The system uses specific nodes such as call (for tool execution) and judge (for decision-making).
  • Control Flow: Workflows support nested calls to other workflows, parallel mapping of tasks, and bounded loops.
  • Cancellations and Recovery: The framework provides host integration for managing execution permissions, cancellation, and state recovery.
  • Integration: The system is split into three primary packages: @parcha/agentrun-dsl for execution and validation, @parcha/agentrun-jev for typed decisions, and @parcha/agentrun-pi for the agent runner extension.

Jev-Powered Decision Making

AgentRun integrates with Jev to implement "typed decisions." This allows workflows to move beyond simple LLM prompts by enforcing confidence thresholds. For example, in a support workflow, a decision node can require a yes answer with a confidence score of at least 0.8 before proceeding. If the confidence is not met, the workflow can be programmed to route the agent to a further investigation step or escalate the task for human review.

Implementation and Use Cases

Support Workflow Example

In a typical support scenario, AgentRun manages a sequence of search and decision nodes. The workflow searches for an answer using a tool, uses Jev to judge if the answer is fit for the purpose, and if it fails the confidence check, it triggers an agent to investigate further. If the second check fails, the case is escalated for review.

Research Workflow Example

The research demo demonstrates the ability to plan sub-questions, research them in parallel, screen evidence, and synthesize a final report. This demonstrates the system's capacity for complex, multi-step reasoning tasks that require evidence selection and report writing.

Integration with Pi

AgentRun can be integrated into the Pi environment (version 0.87.0+). By installing the @parcha/agentrun-pi extension, users can use commands such as /agentrun demo to load scripted examples or /agentrun run to execute them. Pi can also be used to build new workflows by providing it with natural language instructions, which the agent then converts into a formal AgentRun workflow definition.

Community Perspectives and Technical Trade-offs

Technical discussions around the project have highlighted several considerations regarding workflow engines for agents:

"Graph workflows are useless if there are edges without cost or effort attributes and termination states without reward... I thought some time ago about expressing SDLC as combination of harness tools, finite state machines with weights... The goal - model queue of tasks as sub-workflows and implement the most valuable."

Other contributors have noted that while AgentRun provides a specialized DSL, the industry already has durable workflow solutions like Temporal, Erandril, or Airflow, though AgentRun's specific focus is on the agentic loop and the integration of typed decisions.

One developer noted that their own approach to workflow definitions prefers using a "real" language like TypeScript directly to ensure both the user and the agent can understand the function without needing a custom DSL, deriving the UI from the AST (Abstract Syntax Tree) of the code.

Sources

Related

  • Project
    AgentWorkforce/relay

    Agent Relay is an open‑source toolkit that lets teams run LLM‑based coding agents together with existing tools (GitHub, Slack, Notion, etc.). It provides messaging, virtual‑filesystem integrations, searchable session logs, and TypeScript‑defined deterministic flows, enabling auditable, rule‑driven automation of code‑generation and review tasks.

  • Project
    Agent-Field/agentfield

    AgentField is an open‑source Go control‑plane plus Python/Go/TS SDKs that turn ordinary functions into production‑ready, auto‑scalable AI agents reachable via REST. It handles routing, retries, async execution, tracing, cryptographic identity, and provides a web UI for DAG‑style observability.

  • Project
    agentspan-ai/agentspan

    A durable runtime for AI agents that prevents execution loss during process crashes and supports long-running, event-driven workflows with human-in-the-loop approvals.

  • Project
    codeany-ai/open-agent-sdk-typescript

    An open-source TypeScript SDK for running autonomous AI agents in-process, featuring a vast library of built-in tools and support for multiple LLM providers.

  • Project
    agent-sh/agentsys

    AgentSys is an npm‑installable, modular runtime that orchestrates AI agents for end‑to‑end software‑development pipelines. It provides a marketplace of plugins (agents, skills, commands), a linter for agent configs (agnix), and a set of 24 CLI commands (e.g., `/next-task`, `/prepare-delivery`) that compose agents into structured workflows. The system works with Claude Code, OpenAI/Anthropic‑compatible APIs, Codex CLI, Cursor, and Kiro, and includes benchmarks showing cheaper models achieve the same quality when used with its pipelines.