Agents
Configure an agent, don't code one
An agent is a system prompt, a model, your knowledge and your tools, set up in the dashboard and answering through the API, the widget or a workflow. Change it there and every channel picks it up within seconds.
200 free credits when you sign up. No card needed.
What an agent is
Each agent has its own page with eight tabs. Everything the agent knows and does is set there; nothing is hidden in code.
| Settings | Name, description, model and provider, temperature (0 to 2), max tokens, output format (text, markdown, JSON or code), top-p, frequency and presence penalties, and the Active switch |
|---|---|
| Prompt | The system prompt the agent answers with |
| Knowledge | PDFs, websites, snippets and Q&A it searches before answering (how search works) |
| Tools | MCP servers it can call mid-answer (MCP) |
| Deploy | The agent id for the API, the widget's embed code and publishable key, and its allowed websites |
| Conversations | Every conversation from every channel |
| Versions | A history of settings and prompt changes, with a diff and Apply this version |
| Cost | Tokens and estimated model cost, from the cost per million tokens you set |
Models and keys
Agents run on OpenAI, Anthropic, Google Gemini or Groq, on your own provider key. You pay the provider for tokens at their prices; SentientOne doesn't mark them up.
- Save keys once on the LLM keys page and pick one per agent. Keys are encrypted at rest (AES-256-GCM) and never sent back to the browser; the dashboard shows them masked.
- Current models from each provider are in the model list, from small fast ones to the largest reasoning models.
- Need a different key for one call? Send it in the
X-Llm-Keyheader and it is used for that request only. - Switch an agent's provider or model in Settings. Your integration doesn't change: the request and the stream stay the same.
Playground
Chat with any agent before you ship it. Each reply shows its latency, tokens, the tools it called and the sources it used.
- Edit the system prompt as a draft beside the chat. While the draft differs from the live prompt, your messages use the draft, so you can try a change without touching production.
- Publish the draft when it's right; it becomes the agent's prompt and a new version.
- Playground replies cost credits like any other reply (1 each) and show up in Tracing.
The same trick works from code: the API takes a prompt field that replaces the system prompt for one request.
{ "message": "Summarise our refund policy in one line.", "prompt": "You are a terse support agent. Answer in one sentence."}Prompts and versions
Every saved change to an agent's settings or prompt writes a version: who changed it, when, and a line-by-line diff against the one before.
- Apply this version restores an older snapshot and records the restore as a new version, so the history only grows.
- Agents can be switched off with Active. An inactive agent refuses API and widget requests (403 AGENT_INACTIVE) until you switch it back on.
Grounding and guardrails
An agent that uses knowledge gets a grounding guardrail added to its instructions automatically, so it answers from your content rather than from what the model remembers.
For checks on input or output, such as PII, prompt injection, toxicity, off-topic questions, profanity or your own rule, put the agent in a workflow with a Guardrail block.
Deploy
One agent, three ways out. Each uses the same prompt, model, knowledge, tools and credits, and each reply is traced.
| Streaming API | POST /v1/chat/stream from your server with a secret sk-so- key |
|---|---|
| Website widget | One script tag with a publishable pk-so- key, limited to your websites |
| Workflows | An Agent step inside a multi-step automation, run by webhook or API |
Limits
| Agents per account | Unlimited |
|---|---|
| Requests per minute | 60 per agent by default (HTTP 429 RATE_LIMIT_EXCEEDED above it) |
| Message length | 32,000 characters |
| Reply time limit | 120 seconds, then a TIMEOUT error event |
| Cost | 1 credit per reply; a failed reply is refunded |
Build your first agent in the next ten minutes
Sign up and you get 200 free credits, no card needed. Create an agent, add your knowledge, try it in the Playground, then call it from your code or put it on your website.