Description
AgentRun is an open-source platform developed by Parcha Labs that significantly enhances the capabilities of AI agents. It provides a robust control plane, enabling agents to learn from their experiences, execute tasks durably, and manage their operations more efficiently. The platform integrates seamlessly with popular agent harnesses such as Claude Code, Codex, Pi, and OpenCode, offering a unified environment for agent management.
Read More
One of AgentRun's core features is self-learning. Agents become smarter and more cost-effective with each run by identifying repeated steps and converting them into reusable skills and tested workflows. These proposals are validated and graded before acceptance, ensuring that only effective learning is integrated. This process reduces token usage by allowing agents to reuse established methods instead of repeatedly planning and discovering tools, leading to significant cost savings, with some reported cases showing up to a 90% reduction in compliance workloads.
The platform offers a flexible architecture, allowing users to choose their preferred agent harness while AgentRun manages runs, workspaces, recovery, and learning through a single API. It supports various compute environments, from lightweight sandboxes for quick scripts to development environments for building and testing code, and high-resource sandboxes for larger datasets. This separation of the agent loop from the compute environment ensures optimal resource allocation and consistent API interaction across different providers like Docker, Daytona, and Modal.
Durable execution is another critical aspect, ensuring that agent progress is preserved even if a run is interrupted. AgentRun saves conversations, files, and completed tool results, allowing runs to recover from saved states without rebuilding context. This recovery layer is consistent across different harnesses, providing reliable persistence for ongoing tasks. The platform also includes comprehensive agent management features, offering a console and API to monitor agent activities, inspect results, and review learned notes.
AgentRun is available as an Apache-2.0 open-source project, allowing users to self-host it on their own infrastructure with their model API keys, maintaining full control over agents and data. For those preferring a managed solution, AgentRun Cloud provides a hosted deployment managed by the Grep team, eliminating the need for server maintenance. Enterprise options are also available for tailored deployments with dedicated support.
- CategoryAI Agent Builders
- Pricing ModelOpen Source
- DeploymentSelf-Hosted, Cloud
- PlatformsWeb
AgentRun's Core Features
Self-Learning Agents: Agents learn from runs, creating reusable skills and workflows.
Durable Execution: Recovers interrupted runs from saved state, preserving progress.
Shared Control Plane: Manages agents, runs, and learning across multiple harnesses.
Compute Environment Selection: Choose from lightweight, development, or high-resource sandboxes.
Harness Compatibility: Works with Claude Code, Codex, Pi, and OpenCode.
Token Usage Reduction: Reuses learned skills to minimize repeated planning and tool discovery.
Agent Management Console: Monitor agent activities, inspect results, and review learning notes.
Open-Source Platform: Apache-2.0 licensed for self-hosting and full control.
Workflow Authoring: Build inspectable and testable workflows combining code, tools, and agents.
API Access: Common API for creating agents, starting runs, and inspecting results.
Version Control for Agents: Manage agent definitions and revisions.
File Management: Each run gets its own files; upload inputs and inspect outputs.
Hosted Deployment Option: AgentRun Cloud for managed infrastructure.
Enterprise Support: Tailored deployments and support for large organizations.
How to use AgentRun?
Install Client: Install the AgentRun client using npm or pip.
Configure Environment: Set AGENTRUN_BASE_URL and AGENTRUN_API_KEY for your deployment.
Define Agent: Create an agent definition, specifying its harness, instructions, and tools.
Enable Learning: Configure agents to enable self-learning for skill and workflow generation.
Choose Compute: Select the appropriate compute provider (Docker, Daytona, Modal) for agent execution.
Start Runs: Initiate agent runs through the common API or console.
Monitor & Inspect: Use the console to monitor progress, inspect results, and review learning notes.
Reuse Skills: Agents automatically reuse accepted skills and workflows in subsequent runs.
AgentRun's Use Cases
- Automated Compliance Workloads
- Software Development & Testing
- Data Processing & Reporting
- Customer Support Automation
- Research and Analysis
- Workflow Automation
- Agent Experimentation
- Multi-Agent Orchestration




