AI Agents

Agents that plan,
act, and ask for help

Neureus agents run a ReAct loop — Reason → Act → Observe → repeat — until the task is done. Define tools, set a goal, and the agent figures out the steps. Built-in human-in-the-loop approvals when you need a human in the loop.

Start free API reference

ReActloop architecture

HITLhuman approvals

Fullaudit log

AnyLLM model

How the ReAct loop works

The agent iterates until it reaches a final answer. At each step it decides whether to call a tool or return a result.

💭

Reason

Given the goal and tool results so far, what should I do next?

→

⚡

Act

Call a tool: web search, database query, API call, code execution

→

👁

Observe

Tool returns result; agent updates its understanding

↩

Build an agent in 3 API calls

Create the agent

curl -X POST https://app.neureus.ai/agents \
  -H "Authorization: Bearer $NEUREUS_KEY" \
  -d '{
    "name": "Research agent",
    "model": "claude-sonnet-4-6",
    "system": "You are a research assistant. Use tools to find accurate information.",
    "tools": [
      {
        "name": "web_search",
        "description": "Search the web for current information",
        "parameters": {
          "type": "object",
          "properties": {
            "query": { "type": "string" }
          }
        },
        "endpoint": "https://api.yourapp.com/tools/search"
      }
    ]
  }'

Run the agent

curl -X POST https://app.neureus.ai/agents/{id}/run \
  -H "Authorization: Bearer $NEUREUS_KEY" \
  -d '{
    "goal": "Research the current state of AI agent frameworks and summarize the top 5 options with pros and cons"
  }'

Poll for result

curl https://app.neureus.ai/agents/{id}/runs/{runId} \
  -H "Authorization: Bearer $NEUREUS_KEY"

// Response:
{
  "status": "completed",
  "result": "...",
  "steps": [
    { "type": "thought", "content": "I should search for recent agent frameworks..." },
    { "type": "tool_call", "tool": "web_search", "args": { "query": "..." } },
    { "type": "observation", "content": "..." },
    ...
  ],
  "totalSteps": 7,
  "model": "claude-sonnet-4-6"
}

Human-in-the-loop approvals

Some agent actions need human sign-off. Configure approval steps into your agent's workflow — the agent pauses until a human approves or rejects.

Agent reaches approval step

↓

Status: waiting_approval

↓

Your app notifies reviewer (email/Slack)

↓

Reviewer approves via dashboard or API

↓

Agent continues execution

// Add approval step to a workflow
{
  "steps": [
    { "type": "agent_run", "agentId": "...", "goal": "Draft a contract" },
    {
      "type": "human_approval",
      "title": "Review contract draft",
      "assigneeRole": "admin",
      "timeoutHours": 24
    },
    { "type": "agent_run", "agentId": "...", "goal": "Send approved contract" }
  ]
}

// Approve via API
POST /workflows/approvals/{approvalId}/decide
{ "decision": "approved", "note": "LGTM" }

// Or let the agent check its own approval status
GET /agents/{id}/runs/{runId}
// { "status": "waiting_approval", "approvalId": "..." }

What's included

🔁

ReAct loop

Iterative reasoning with tool calls. Agent stops when it has a final answer or reaches the max step limit.

🔧

Custom tools

Define tools as HTTP endpoints. The agent calls them with structured JSON arguments and incorporates the response.

🧠

Any model

Claude, GPT-4o, Llama, DeepSeek — configure per-agent. Mix models across agents in the same workflow.

👤

HITL approvals

Pause agent execution at configurable steps. Reviewers approve or reject via API or the Neureus dashboard.

📋

Audit log

Every thought, tool call, and observation is logged with timestamps. Full run history for debugging and compliance.

🏪

Marketplace

Publish agents to the Neureus marketplace. Other tenants can install and run your agents without seeing your implementation.

Agents that plan,act, and ask for help