@molecule/api-ai-agents-llm

Auto-generated, AI-first package reference for the molecule.dev ecosystem. It is written to be read by coding agents as much as by people, and is generated from this package's source — edit src/index.ts JSDoc, not this file.

LLM-backed AI agents provider for molecule.dev — the batteries-included tool-calling loop that implements the @molecule/api-ai-agents core contract.

Ships a default provider (name: 'llm') that drives a model↔tool loop over the swappable ai chat bond (@molecule/api-ai): it calls the model, executes any tools the model requests via each AITool.execute(), feeds the results back, and repeats until the model stops calling tools (or the maxSteps budget is exhausted). Bond it once at startup, then drive it from anywhere via the @molecule/api-ai-agents accessor.

Quick Start

import { bond } from '@molecule/api-bond'
import { requireProvider } from '@molecule/api-ai-agents'
import { provider as agents } from '@molecule/api-ai-agents-llm'
import type { AITool } from '@molecule/api-ai'

// Wire at startup (an `ai` provider must already be bonded).
bond('ai-agents', agents)

const myTool: AITool = {
  name: 'add',
  description: 'Add two numbers',
  parameters: {
    type: 'object',
    properties: { a: { type: 'number' }, b: { type: 'number' } },
    required: ['a', 'b'],
  },
  execute: async (input) => {
    const { a, b } = input as { a: number; b: number }
    return a + b
  },
}

const result = await requireProvider().run({
  task: 'What is 1 + 2? Use the add tool.',
  tools: [myTool],
})
console.log(result.output) // final assistant answer
console.log(result.steps) // intermediate tool calls + results
console.log(result.usage) // token usage summed across all model calls

Type

provider

Installation

npm install @molecule/api-ai-agents-llm @molecule/api-ai @molecule/api-ai-agents @molecule/api-i18n

API

Classes

AgentRunError

Thrown by run() when the agentic loop fails partway through — a provider API error (rate limit, auth, overload, network), an abort, or any other exception raised while draining a model turn or executing a tool.

Constants

provider

The LLM-backed AI agents provider: a tool-calling loop over the bonded ai chat provider.

Bond it with bond('ai-agents', provider) and drive it with requireProvider().run({ task, tools }). Requires a bonded ai provider whose model supports tool use.

const provider: AIAgentsProvider

Core Interface

Implements @molecule/api-ai-agents interface.

Bond Wiring

Setup function to register this provider with the core interface:

import { setProvider } from '@molecule/api-ai-agents'
import { provider } from '@molecule/api-ai-agents-llm'

export function setupAiAgentsLlm(): void {
  setProvider(provider)
}

Injection Notes

Requirements

Peer dependencies:

• @molecule/api-ai ^1.0.1
• @molecule/api-ai-agents ^1.0.1
• @molecule/api-i18n ^1.0.1

Runtime Dependencies

• @molecule/api-ai
• @molecule/api-ai-agents
• @molecule/api-i18n

Requires a bonded ai provider (bond('ai', provider) or a named provider selected via run({ provider: 'anthropic' })) whose model supports tool use — this agent has no model of its own; it orchestrates the ai bond.

The loop executes your tools: each model tool_use runs the matching AITool.execute(input) and the return value is fed back as a tool_result. A thrown tool error or an unknown tool name becomes an error tool_result (recorded with isError: true) instead of aborting the run, so the model can recover. run() requires exactly one of task (a single user turn) or messages (a full history); it throws if neither — or both — is supplied.

Every model turn streams by default (input.stream — set false to force non-streaming) and, when input.cacheControl is set, is passed through to ai.chat() on every turn so a provider that supports prompt caching (e.g. @molecule/api-ai-anthropic) doesn't re-bill the identical system + tools prefix on each of up to maxSteps turns. Note onEvent is only an INCREMENTAL live hook while streaming — with stream: false it fires once per turn with the whole response.

Failure semantics (how to tell failure modes apart):

• Tool failure (throwing execute(), unknown tool name) — NON-fatal: recorded on the step with isError: true and fed back to the model.
• AI provider API failure (rate limit, bad key, overload, network — the ai bond emits an in-band error ChatEvent), an abort, or any other mid-run exception — FATAL: run() rejects with an AgentRunError (message text unchanged from a plain Error) carrying the usage and steps accumulated across every turn that completed before the failure — a caller that meters spend can book that partial usage instead of losing it. run() never resolves with a silently empty output, so an empty result.output means the model genuinely produced no text, not that the API failed.
• Step budget exhausted — resolves normally; output is the last assistant text or an explicit "step budget exhausted" note.

Swappable like any bond: replace this LLM agent with your own AIAgentsProvider via bond('ai-agents', myProvider) — nothing else changes.

E2E Tests

Integration checklist — drive the real UI (live preview, no mocks), give the agent a task that genuinely needs several tool calls, adapt each item to this app's actual agent surface + its registered tools, and check every box off one by one. A box you can't check is an integration bug to fix — not a skip:

• A multi-step task drives a real agentic LOOP, not one shot: run() records more than one entry in result.steps, and each toolCalls[] names a tool the agent was actually given. Confirm each tool's execute() truly RAN via its own side effect / log / the resource it touched — not that the model narrated calling it — and that result.output is a final answer that USES those tool results to complete the task.
• The loop TERMINATES cleanly: it either reaches a done state (the model stops calling tools, so run() returns the final output) or hits the maxSteps cap (default 10) and stops with the step-budget note as output. It never spins forever or ping-pongs the same tool endlessly — a step cap exists and is honored: result.steps.length <= maxSteps, and a task built to loop past the budget stops AT the cap rather than running away.
• A tool that fails mid-loop is handled, not fatal: when a tool's execute() throws (or the model names a tool that was never registered), that call is caught and fed back to the model as an error tool result — the matching toolCalls[] has isError: true, the loop continues, and the run still returns a result. The request never crashes. (Only an ai-provider failure rejects — as AgentRunError carrying the partial usage/steps — and it surfaces as a handled error, not an unhandled throw.)
• If the run streams (the default), the UI shows progress INCREMENTALLY: the onEvent hook fires per ChatEvent (thinking text, tool_use, done) as they arrive, so the user watches the agent think + call tools live — not a long freeze then one blob at the end.
• SECURITY — the agent can call ONLY the tools handed to this run: a tool name the model invents (or one injected via the task or a tool's own output) that was never in input.tools is refused as an unknown tool error, not executed. Tool execution stays server-side under its backend's guards (e.g. the @molecule/api-ai-tools pathGuards/redactSecrets/blockCommand), and the ai provider key lives on the API, never the client. Feed a prompt-injected instruction (in the task, or in data a tool returns) telling the agent to escape the workspace, run a privileged command, or exfiltrate a secret, and confirm it CANNOT do anything the user couldn't do directly.