SandBase AI has published an expanded version of its Awesome Agent Runtime repository, cataloging 500 projects across ten distinct infrastructure layers that production AI agent systems require. The map moves beyond the early agent conversation—prompts, tools, demos—and tackles the unglamorous work of tracking where agents actually run in real products: code execution environments, browser automation primitives, tool protocols, memory layers, observability stacks, and deployment compute. The repository went live on DEV.to on June 26, 2026.
Why Agent Infrastructure Deserves Its Own Map
The early agent discourse focused heavily on prompting techniques and demo showcases. That's useful for understanding what's possible in a sandbox environment, but production systems face a different reality. Once agents touch real files, browsers, APIs, credentials, workflows, and customer data, the infrastructure layer becomes the actual product boundary. Teams need to answer hard questions: where does code run? What can tools access? How do you recover when an agent behaves unexpectedly? SandBase built this map because those questions deserve structured answers.
Inside the 10-Layer Stack
The repository organizes projects across ten categories that represent the full production agent stack: Agent Runtime, Execution Sandbox, Browser Automation, Tool Protocol, App Integrations, Memory and Context, Safety and Evals, Model Gateway, Observability, and Deployment or Compute. Each category captures infrastructure for controlling state and workflows, running code safely, automating browsers, connecting tools to applications, storing memory and context across sessions, evaluating agent behavior, routing between models, observing failures with traces and cost visibility, and deploying workloads at scale.
What the Ecosystem Map Reveals
Several patterns emerged from curating the first 500 projects. Agent frameworks are maturing quickly, but runtime safety remains uneven—builders increasingly ask not just whether an agent can call a tool, but where that tool runs, what it can access, and how to recover when things go wrong. Execution sandboxes and browser automation are becoming first-class primitives rather than afterthoughts. MCP and tool protocols like Model Context Protocol are giving the ecosystem shared language for how agents interact with tools, permissions, and app integrations. Observability is moving beyond basic prompt logs toward traces, evals, cost visibility, tool-call history, runtime events, and failure analysis. Deployment is also fragmenting—model inference represents only one slice of a stack that needs separate operating models for sandbox execution, tool infrastructure, browser workers, and integration runtimes.
How Builders Can Use the Repository
The map serves as a comparison framework across infrastructure categories, a discovery mechanism for projects to integrate, a diagnostic tool for identifying gaps in an existing agent stack, and a networking surface for finding maintainers and adjacent ecosystems. SandBase is actively asking contributors who maintain or use relevant projects to open issues or pull requests with additions covering agent runtimes, execution sandboxes, browser automation infrastructure, MCP implementations, memory layers, evals, guardrails, observability platforms, model gateways, and deployment compute.
The Bottom Line
This map matters because it codifies what experienced builders already knew: agents in production need real infrastructure, not just better prompting. If you're shipping agent-based products and haven't thought through your sandbox strategy, your eval framework, or how you'll observe failures at runtime—you're flying blind. Bookmark the repo. You'll need it.