From isolated bots to an ecosystem of open source AI agent frameworks
Open source AI agent frameworks are software toolkits and standards that define how AI agents discover tools, write backend code, run safely in production, and expose transparent telemetry across local and cloud environments, turning isolated bots into connected, observable systems that can scale with standardized patterns instead of one-off integrations. Today’s agent developers face four recurring problems: how agents find tools, how they build production backends, how teams control costs and “shadow AI,” and how leaders gain visibility into what agents do on real machines. A new wave of projects is tackling these gaps from different angles: Google and Microsoft’s Agentic Resource Discovery (ARD) specification, Vercel’s eve framework with Passport governance, AWS Blocks for AI agent backend development, and Agent Beacon for agent telemetry tools. Together they signal a move toward a shared AI tool discovery specification, predictable infrastructure, and traceable behavior.
Agentic Resource Discovery: a shared AI tool discovery specification
Agentic Resource Discovery (ARD) focuses on the earliest step in any workflow: how agents find and verify tools at runtime. Instead of wiring every API, MCP server, or peer agent ahead of time, ARD turns discovery into a search request across catalogs and registries. Each publisher hosts an ai-catalog.json on its own domain, listing its tools and agents, while registries crawl those catalogs, index them, and respond to natural-language discovery queries. The spec uses domain ownership and optional cryptographic trust metadata so agents can confirm who published a capability before connecting. According to Search Engine Journal, eleven companies including Google, Microsoft, GitHub, and Hugging Face backed the draft ARD spec under an Apache 2.0 license, building on an AI Catalog data model maintained under the Linux Foundation. For open source AI agent frameworks, ARD promises a standard way to plug into a growing web of capabilities.
Vercel eve: opinionated agents, cost controls, and shadow AI governance
Vercel’s eve framework tackles the operational side of AI agent backend development: predictable behavior, lower costs, and control over employee-built agents. In eve, an agent is a directory containing TypeScript and Markdown files that define instructions, skills, tools, model providers, authentication, channels, and schedules. Agents run in sandboxed virtual machines by default, and a built-in testing tool exercises them and evaluates results. The framework integrates with Vercel’s AI SDK and can route calls through the AI Gateway to consolidate traffic and optimize spending. Passport, announced alongside eve, aims to pull “shadow AI” projects under enterprise governance by tying these agents to organizational policies and identity. Eve’s design favors a “fill in the blanks” style over heavy configuration, making it attractive for teams that want standardized, open source AI agent frameworks without exposing infrastructure details to every developer or agent builder.
AWS Blocks: AI agents that can write production-grade backends
AWS Blocks concentrates on making AI agents first-class backend developers. It is an open-source TypeScript framework where each Block bundles application code, a local development version, and the AWS infrastructure needed in production. Developers can run a single development command and get local Postgres, authentication, real-time messaging, and file storage without an AWS account, then deploy the same code to Lambda, DynamoDB, Aurora, API Gateway, and Bedrock with no changes. What sets Blocks apart is that it ships with steering files that guide coding agents toward the right architecture and patterns. An instruction like “add authentication and a database” can yield code that works locally and promotes cleanly to production services. Blocks model capabilities such as databases, user auth, AI agents, file uploads, and background jobs as composable npm packages, helping agents generate type-safe, deployable backend code instead of fragile prototypes.
Agent Beacon: telemetry and observability for local AI coding agents
Agent Beacon addresses a quieter but critical problem: knowing what AI coding agents are doing on real machines. Developed by Asymptote Labs, Beacon is an open-source telemetry layer that runs locally and records normalized event logs for tools like Claude Code, Codex CLI, Cursor, and Claude Cowork. It discovers supported runtimes, configures them to export OpenTelemetry data, and installs hooks where needed to capture sessions, prompts, tool use, command execution, MCP-like tool activity, approval decisions, and file edits. A bundled collector converts these signals into JSON events, which feed a local dashboard or can be forwarded into customer-managed SIEMs. Written mostly in Go and installable via Homebrew, Beacon offers flexible retention modes, from full telemetry to redacted logs that strip prompt text or command output. For teams experimenting with AI agent backend development, Beacon’s agent telemetry tools bring the transparency and audit trail that production workflows demand.
