@latimer-woods-tech/agent

Cloudflare-native LLM agent runtime for the Factory platform. One hardened orchestration engine — tool registry, reasoning loop, memory, guardrails — that powers the vertical SaaS products (Voice, Video, Astrology).

See docs/architecture/AGENT_RUNTIME.md for the full design.

Status

Incremental build (Agent Runtime plan, Phase 2):

• ✅ Tool registry — the shared seam between the deterministic supervisor planner and the LLM reasoning loop.
• ✅ Reasoning loop — tier/scope/schema enforcement, quarantined receipts, provider-aware budgets, and explicit failure states.
• ⏳ Memory tiers (working / episodic / semantic)
• ⏳ Agent recipes + gateway

Tool registry

import { ToolRegistry, type Tool } from '@latimer-woods-tech/agent';

const registry = new ToolRegistry();
registry.register({
  name: 'humandesign.read.blueprint',
  description: 'Fetch a blueprint by id',
  side_effects: 'read-external',
  required_scope: 'read',
  parameters: { type: 'object', properties: { id: { type: 'string' } }, required: ['id'], additionalProperties: false },
  invoke: async (slots) => ({ ok: true, result: await fetchBlueprint(slots.id) }),
});

// Deterministic planner path — match by name, filter by trust tier.
registry.byTier('green');

// LLM path — only schema-bearing, opted-in tools, optionally tier-scoped.
const tools = registry.llmTools('green'); // → pass to @lwt/llm complete({ tools })

A tool with parameters (JSON Schema) is offered to the LLM by default; omit parameters for a template-only tool, or set exposeToLLM: false to keep a schema'd tool off the LLM surface.

Every loop invocation supplies controller-owned grantedScopes and may narrow the registry further with allowedTools. Mutations additionally require a controller-issued runId and durable journal; completed calls replay their sanitized receipts, while pending/unknown outcomes escalate without reinvocation.