Felix: Self-Hostable Managed Agents Harness

What changed

Felix has been released as a self-hostable managed agents harness, enabling developers to define and manage AI agents through YAML manifests. The system compiles these manifests into governed agents that incorporate durable fibers for execution, memory capabilities, skill integration, evaluation tools, approval workflows, and sandboxing. Felix supports multiple LLM backends, including OpenAI, A2A, MCP, and SSE, and offers various communication surfaces such as REST/SSE, an OpenAI-compatible API, A2A, and MCP.

Key features include:

• Manifests: Agents are defined using felix/v1 YAML, with bundled agents available in the manifests/ directory.
• Governance: Built-in features for authentication, approvals, auditing, and usage metering.
• Durable Execution: Utilizes fibers for reliable execution, with optional Temporal integration, allowing for steering and follow-up actions.
• Session Control: Supports forking, rewinding, and managing session history with compacting, windowed, or semantic strategies.
• Memory and Skills: Provides durable facts, procedural memory, and Agent Skills.
• Deployment: Designed for flexible deployment using Docker, Helm, AWS, or GCP, with options for lean images suitable for small VMs.
• Observability: Integrates with Prometheus for metrics and supports OpenTelemetry for tracing and logging, with options for deploying a full observability stack.

The quick start guide demonstrates setting up Felix locally using Docker Compose, including database migrations and starting the API and worker services. It also shows how to send requests via the chat API or the OpenAI-compatible endpoint, using a local API key generated during setup. For local development without Docker Compose, commands are provided to install dependencies, run migrations, and start a development server.

Deployment options cater to different needs, from lean Docker images for resource-constrained environments to full configurations including cloud storage integrations for AWS and GCP. Observability is a key aspect, with Prometheus metrics available by default and optional OpenTelemetry integration for detailed tracing and logging. An analytics warehouse feature is also available, supporting DuckDB, ClickHouse, or Doris for audit and evaluation data.

Why it matters for builders

Felix empowers AI builders by providing a structured and governed environment for developing and deploying AI agents. The YAML manifest system simplifies the definition of agent behavior, making it easier to version control and manage complex agent configurations. Support for multiple LLM providers and deployment targets ensures flexibility and avoids vendor lock-in. The inclusion of features like durable execution, memory, and governance tools allows for the creation of more sophisticated and reliable AI applications.

Practical impact

Developers can leverage Felix to build and deploy agents on infrastructure they manage, offering greater control over data and operations. The system's modular design and deployment flexibility mean agents can be scaled from small VMs to cloud environments. Builders can experiment with different LLM backends and agent configurations using the manifest system, and integrate Felix into existing CI/CD pipelines for automated testing and deployment. The observability features enable monitoring and debugging of agent performance in production.

Caveats and source limits

The provided source is a GitHub repository description and README, offering detailed technical information about Felix's features and architecture. However, it does not include specific benchmark results, pricing information, or a definitive release date beyond the latest version tag (v0.4.0). Information regarding community adoption metrics such as stars and forks is present but limited. The source focuses on technical implementation and deployment, with less emphasis on user-facing tutorials or advanced use cases beyond the quick start.