Reactive Agents — Framework Orientation

Agent objective

After loading this skill you know: (1) what the framework does, (2) the canonical builder chain pattern, (3) which capability skills to load for the task at hand.

Framework overview

Reactive Agents is an Effect-TS layered runtime for building autonomous AI agents in TypeScript. Agents are composed via a fluent ReactiveAgentBuilder — each .withX() call wires in an optional capability layer. The runtime ships 25 packages covering reasoning, memory, tools, MCP, guardrails, identity, observability, orchestration, cost, verification, eval, A2A networking, and web framework integrations (React, Vue, Svelte).

Six LLM providers are supported: anthropic, openai, gemini, ollama, litellm, and test.

Two syntaxes, one API

createAgent(config) (declarative) and ReactiveAgents.create().withX() (fluent) are the same API — same key names, same nesting, validated against the same AgentConfigSchema. Prefer createAgent for static definitions (the 90% case); reach for the builder when construction is conditional/imperative or needs a code-only escape hatch (.withHook, .withLayers, .compose).

Canonical pattern — createAgent(config) (front door)

import { createAgent } from "@reactive-agents/runtime";

const agent = await createAgent({
  name: "my-agent",
  provider: "anthropic",              // required
  model: "claude-sonnet-4-6",         // optional — provider default if omitted
  reasoning: { defaultStrategy: "adaptive", maxIterations: 10 },
  tools: {},                          // enables built-in tools
  memory: { tier: "enhanced", dbPath: "./agent.db" },
  observability: { verbosity: "normal", live: true },
});

const result = await agent.run("Your task here");
console.log(result.output);
console.log(result.metadata.stepsCount, result.metadata.strategyUsed);

Same agent — fluent builder

import { ReactiveAgents } from "@reactive-agents/runtime";

const agent = await ReactiveAgents.create()
  .withName("my-agent")
  .withProvider("anthropic")
  .withModel("claude-sonnet-4-6")
  .withReasoning({ defaultStrategy: "adaptive", maxIterations: 10 })
  .withTools()
  .withMemory({ tier: "enhanced", dbPath: "./agent.db" })
  .withObservability({ verbosity: "normal", live: true })
  .build();                           // always await — returns Promise<ReactiveAgent>

Skill routing — select by what you're building

Building...	Load these skills
Any agent (start here)	builder-api-reference
Task agent (research, analysis, coding)	reasoning-strategy-selection, tool-creation
Agent that runs shell commands	shell-execution-sandbox
Agent with persistent memory	memory-patterns, context-and-continuity
Agent using MCP tools	mcp-tool-integration
Always-on / scheduled agent	gateway-persistent-agents
Multi-agent workflow	multi-agent-orchestration
Agent embedded in a web app	ui-integration, interaction-autonomy
Production / multi-tenant agent	identity-and-guardrails, cost-budget-enforcement
Agent with output quality guarantees	reasoning-strategy-selection, quality-assurance
Agent-to-agent networking	a2a-agent-networking
Custom provider behavior or local models	provider-patterns

Recipe skills — complete reference implementations

Load a recipe skill for a full working example:

Recipe	What it builds
recipe-research-agent	Research/analysis agent with memory + verification
recipe-code-assistant	Code generation + sandboxed shell execution
recipe-persistent-monitor	Always-on monitoring via gateway + crons
recipe-orchestrated-workflow	Multi-agent pipeline, lead/worker pattern
recipe-saas-agent	Multi-tenant agent with identity + cost controls
recipe-embedded-app-agent	Agent in React/Vue/Svelte with streaming UI

Builder factory methods

createAgent(config)                          // declarative front door — validate + build in one call
ReactiveAgents.create()                      // blank builder
ReactiveAgents.fromConfig(config)            // from AgentConfig object → builder
ReactiveAgents.fromJSON(json)                // from JSON string
ReactiveAgents.runOnce("task", builder)      // build + run + dispose in one call
builder.buildEffect()                        // returns Effect<ReactiveAgent> for Effect runtimes

Pitfalls

• .build() is async — always await it; forgetting causes silent "agent is undefined" errors
• .withProvider() is required — there is no default provider
• .withTools() with no args enables 5 standard tools: web-search, http-get, file-read, file-write, code-execute. Shell execution is opt-in only via .withTools({ terminal: true }); use allowedTools to restrict standard tools
• Strategy name is "plan-execute-reflect" — not "plan-execute" (throws StrategyNotFoundError)
• Memory tiers are "standard" and "enhanced" — not "1" and "2" (those are deprecated)
• "groq" and "openrouter" are not valid provider names — use "litellm" for proxy/router providers