Loading AI::COSMOS...
“AI Agents extend LLMs from passive text generators into active problem solvers. The ReAct (Reasoning + Acting) architecture alternates between generating a Thought (internal reasoning), executing an Action (invoking external APIs, SQL databases, or Python code execution via structured JSON function calling), and receiving an Observation (tool return payload) in a closed loop until the task is resolved.”
Orchestrating autonomous AI agents through Reasoning + Action (ReAct) loops, JSON schema tool calling, and multi-agent delegation.
// ReAct Autonomous Agent Execution Loop
export interface Tool {
name: string;
description: string;
execute(args: Record<string, any>): Promise<string>;
}
export async function runAgentLoop(
task: string,
tools: Tool[],
maxSteps = 5
): Promise<string> {
const history: string[] = [`Task: ${task}`];
for (let step = 0; step < maxSteps; step++) {
const prompt = history.join('\n') + '\nThought:';
const response = await callLLM(prompt); // e.g. "Thought: Need to query SQL -> Action: sqlQuery"
if (response.includes('Final Answer:')) {
return response.split('Final Answer:')[1].trim();
}
const { toolName, args } = parseAction(response);
const tool = tools.find(t => t.name === toolName);
const observation = tool ? await tool.execute(args) : 'Tool not found';
history.push(`Action: ${toolName}(${JSON.stringify(args)})`);
history.push(`Observation: ${observation}`);
}
return 'Max execution steps reached.';
}Goal Input: User assigns complex multi-step objective
Thought Generation: LLM plans intermediate sub-goal and determines required tool
Action Execution: LLM outputs valid JSON function call with typed arguments
Tool Execution: Host environment executes API/database call and captures result
Observation Loop: Tool output fed back into LLM context; repeats until Final Answer generated
Strict JSON Schema validation (OpenAI Structured Outputs / Pydantic) guarantees 100% schema conformance for downstream function calls.