beam_agents

beam_agents defines what an agent is and ships a tool-loop agent that talks to any OpenAI-compatible model.

beam_agents = { path = "../riverbeam/crates/beam_agents" }

Types

Item	Role
`Agent`	Trait: `metadata()`, `stream(input, context)`, `health_check()`
`AgentRegistry`	Registered agents: `register`, `get_metadata`, `dispatch_frames`, `dispatch_stream`, `list_agents`
`EchoAgent`	Replies `Echo: <input>`; for tests and demos
`RigAgentRunner`	Tool loop over a completion client: `new(name, client, tools, skill)`, `for_skill(skill, client, tools)`
`CompletionClient`, `Completion`, `PendingToolCall`	The model boundary
`OpenAiCompatible`	Chat completions over HTTP: `new`, `from_env`, `with_max_tokens`, `with_extra_body`
`EchoClient`, `ScriptedClient`	Clients without a network, for development and tests
`Tool`, `ToolSpec`, `json_params`	Tools the model can call
`sandbox_tools(Option<WorkingTree>)`	Nine file and git tools
`RememberTool`	The `remember` tool spec; invoking it returns `BEM-109`
`Skill`, `builtin_skills()`, `skill_by_name()`	System prompts with tool allowlists
`terminal_result`, `human_message_from_input`	Helpers for agent authors

There are no feature flags.

Usage, the tool list, the skills and the current limitations are in Agents, tools and skills.