Today we are releasing Clear Ideas Agent Runtime 0.1.0, a standalone, open-source TypeScript runtime for defining, running, and observing portable AI agents.
AI agents often begin as application code that combines prompts, control flow, model providers, credentials, tools, persistence, and deployment assumptions. That coupling makes an agent harder to inspect and harder to move when the execution environment changes.
Agent Runtime separates the agent definition from the infrastructure that executes it. Install it from npm, inspect the source on GitHub, and run it without a hosted Clear Ideas service.
npm install @clearideas/agent-runtimePortable Agent Definitions
An Agent Manifest is a versioned YAML or TypeScript contract describing steps, variables, conditions, loops, tools, outputs, and limits. An Agent Run Manifest references that definition and supplies validated values for one invocation.
Credentials, model configuration, connections, persistence, compute, sandboxes, and telemetry remain under host control. A manifest can therefore move from a local process into an application, a child-process worker, or remote compute without being rewritten around the new environment.
The manifest documentation covers the two contracts, variable validation, conditions, loops, model selection, structured output, and run-time overrides.
One Runtime Across Execution Boundaries
Agent Runtime supports three execution boundaries through one client contract:
- In process for embedding the runtime directly in a TypeScript application
- Child process for running through the portable worker protocol
- Remote compute through an execution-engine adapter
The run contract does not change when the execution boundary changes. Applications can begin with local execution and later dispatch the same Agent Manifest through another process or remote worker.
The included Modal adapter implements remote submission, status, ordered events, result retrieval, resume, and cancellation. The remote execution guide documents the worker protocol, HTTPS worker, compute-adapter interface, and durable control-plane requirements.
Durable State and Ordered Events
Agent Runtime checkpoints run state, step outputs, transcripts, generated files, and continuation data. Suspended or interrupted runs can resume in a fresh process. Attempt fencing prevents an older worker from committing after a resumed attempt takes ownership. The persistence and recovery guide describes checkpoints, resume, suspension, and reconciliation.
Runs emit ordered lifecycle, model, tool, checkpoint, artifact, and text-delta events. The same stream can drive a CLI, an embedded application, a remote client, or an OpenTelemetry integration. Event cursors include both the execution attempt and sequence, allowing a client to reconnect after a worker has been replaced.
Native Authorization and Isolation
Authorization is part of the runtime boundary. Agent Manifests cannot grant themselves credentials, infrastructure access, or broader tool permissions. The host resolves credentials, configures connections, defines tool allowlists, narrows read and write modes, and can authorize each tool request against application policy. See connections and tools for the MCP and application-tool contracts.
Code execution and generated files use native sandbox contracts. The project includes providers for Docker Engine and Modal Sandboxes, including the Modal JavaScript Sandbox API. Applications can supply another sandbox implementation through the same adapter boundary.
Authorization and sandboxing are not hosted-service features added outside the package. They are runtime contracts used when resolving connections, invoking tools, running code, and projecting generated artifacts.
Models, Tools, Stores, and Adapters
Agent Runtime includes prompt, condition, collection-loop, goal-loop, approval, webhook, code, and sub-run steps. Models connect through the AI SDK, while tools can come from the host application or a Model Context Protocol connection.
Memory, file, and SQLite stores cover common persistence needs. Separate adapters connect models, condition evaluators, compute, sandboxes, generated files, and telemetry, allowing an application to install only the infrastructure it uses.
The adapter catalog lists the packaged integrations and the interfaces for custom infrastructure.
Dependency-Aware Scheduling
Sequential execution remains the default. A run can also ask Agent Runtime to analyze manifest outputs and variable references, resolve eligible prompt work into dependency-aware execution waves, and run independent branches concurrently. Results commit in manifest order, while tool-enabled prompts and other stateful work remain ordered.
This is one runtime capability rather than the definition of the product. It is useful when an agent performs independent research, extraction, classification, or drafting before synthesis.
Read How Agent Runtime Resolves Manifests into Execution Graphs for the complete manifest, dependency analysis, scheduling rules, timing model, and execution visualization.
Public Packages and Documentation
Version 0.1.0 includes the runtime, CLI, worker protocol, examples, schemas, documentation, and composable infrastructure adapters. The packages are public under the @clearideas scope, carry npm provenance, and are licensed under Apache 2.0.
Agent Runtime is intended for developers who need a TypeScript runtime embedded in an application, portable agent definitions across execution environments, durable multi-step runs, or host-controlled models, tools, credentials, sandboxes, and infrastructure.
Agent Runtime and the Clear Ideas Product
Agent Runtime is an independent open-source project. Hosted Clear Ideas uses the same runtime foundation and adds organization policy, governed records, content permissions, user experience, and managed operations. None of those hosted services are required to install or operate the runtime.
Install and Run
Start with the five-minute quickstart, then use the building agents guide or embedding guide for your application.
You can also review the source repository, install @clearideas/agent-runtime from npm, or open the Agent Runtime overview.