AgentRun platform

AgentRun is an open-source platform that supercharges your agents. It adds self-improvement, durable execution and a control plane on top of your favorite agent harness. Your agents become smarter, more fault tolerant and use fewer tokens with every run. It provides a common API and console around your chosen harness, with persistent runs, compute, and learning.

What runs where

Part	Responsibility
Your application, CLI, or console	Create agents, submit work, answer questions, and inspect results.
AgentRun	Keep agent revisions and run state; coordinate tools, compute, recovery, and learning.
Agent harness	Run the reasoning loop. Supported adapters include Pi, Claude Code, Codex, and OpenCode.
Model provider	Serve model requests using your configured credentials.
Compute provider	Run shell and file tools in an available box when the agent needs one.

Harness choice and compute choice are separate. Selecting Codex, for example, does not require your application to adopt a different run API. Selecting an available compute provider changes where the tools execute.

The objects you work with

An agent is a saved definition: instructions, harness, model, tools, and optional learning and compute settings. Editing it creates a revision. A run executes a particular revision with an input and keeps its history, files, result, and usage.

Learning belongs to the agent. Each new run starts with the accepted learning available at its start, so later work can reuse what earlier work established.

Start with the platform

1. Run your first agent from the local console.
2. Create agents and inspect runs through a consistent lifecycle.
3. Configure models and keys for the harnesses you use.
4. Enable self-improvement for repeated tasks.
5. Choose compute and understand recovery.
6. Operate from the control plane or integrate the API.

You do not need to write a workflow, install a harness extension, or implement a host adapter to run your first agent. Those remain options when you need a custom integration.

Source documentation and installed versions

These guides follow the current AgentRun repository. The imported server reference, learning specification, and other source references identify their source revision.

The quickstart uses the published launcher @parcha/agentrun@0.4. Repository changes can arrive before the next package release. On a running server, GET /v2/capabilities reports available harnesses, tools, models, and compute; GET /v2/openapi.json describes that server's routes. Use those responses when building against a particular installation.

See Packages for the distinction between the launcher, server, clients, and lower-level libraries.