# Agent Playground URL: https://discuvr.org/agent-playground/ Last updated: 2026-11-07 Run AI agents that use tools - web search, a calculator, URL fetching - across multiple providers, and watch each step of the loop as it happens. Agent Playground runs the same task through up to eight tool-using agents at once and shows every step: each tool call, its arguments and what came back. Most of what goes wrong with an agent is in the loop rather than the model, so a failure is something you can see instead of guess at. ## What it does - Three tools you switch on per agent: web search (top five DuckDuckGo results), a calculator and a URL fetcher. - Nine provider options: OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, OpenRouter, xAI and any OpenAI-compatible endpoint. - Every tool call and its result is shown in order, with step and token counts for each agent. - Follow-up messages carry the last 10 messages of each agent's conversation. ## Why the trace matters When an agent gets an answer wrong, the useful question is which step went wrong: a bad plan, a tool called with the wrong argument, or a correct result misread afterwards. A final answer alone cannot tell you which, and the three have completely different fixes. ## What each tool can and cannot do Web search returns titles, links and snippets for the top five results, not whole pages. Fetch URL reads up to 512 KB of a public http or https page, strips the HTML and gives the model the first 4,000 characters of text; private, loopback and other internal addresses are refused. The calculator handles arithmetic and math functions such as sqrt and log, up to 500 characters, and refuses huge powers. No tool runs code. ## Limits and where your keys go Each agent stops after 8 steps. Up to 8 agents per run, task and system message up to 20,000 characters each, sign-in required, and 30 runs per 10 minutes per IP address. Keys come from the same Keys page as AI Playground: they stay in your browser's local storage, are sent with each run and are not saved on the server. ## What happens in one step, provider by provider A step is one model call. The agent sends the conversation plus the definitions of the tools you switched on; if the reply asks for tools, every requested call runs in order, the results are added to the conversation and the next step starts. A reply with no tool call is the final answer. The wire format differs by provider: OpenAI-compatible APIs (OpenAI, Groq, Mistral, DeepSeek, OpenRouter, xAI and Custom) return tool_calls and get each result back as a tool message, Anthropic returns tool_use blocks and gets tool_result blocks, and Gemini returns functionCall parts and gets functionResponse parts. The trace shows the same call and result whichever format carried it. ## Questions **Which tools can the agent use?** Web search, a calculator and URL fetching. Every provider gets the same three, and you choose which ones each agent may call. **Do I need my own API keys?** Yes, one for each provider you run, and you pay that provider directly. Keys stay in your browser, are sent with each run and are not saved on the server. **What happens when an agent runs out of steps?** It stops after its eighth model call, even if it was still calling tools, so the card can end without a final answer. That usually means it is looping: narrow the task or switch off the tool it keeps misusing. **Why does the input token count grow with every step?** Each step resends the whole conversation, including every earlier tool result, so a five-step run sends your task five times and an early search result in every later step. The token figures on each card are the totals across all steps, as reported by the provider. **Does a follow-up message remember the earlier searches?** Only as text. A follow-up carries the last 10 messages of that agent's conversation, which are your earlier tasks and the agent's final replies, not its tool calls or their results. If the answer needs the source again, ask the agent to search or fetch it again. **Can I connect a local model such as Ollama or LM Studio?** Only if it is reachable from the internet. Choose Custom (OpenAI-compatible) and enter the full chat completions URL, for example https://your-host/v1/chat/completions. The request is sent from the Discuvr server, not from your browser, and any address that resolves to localhost, a private network or another reserved range is refused, so http://localhost:11434 (Ollama) or http://localhost:1234 (LM Studio) will not work. Put the server behind a public HTTPS address you control, and save a key for Custom on the Keys page even if your server ignores it, because a run without a key stops with an error. For an agent, also pick a local model that supports tool calling. --- Part of Discuvr (https://discuvr.org/): Discuvr is a collection of free, focused tools for day-to-day professional work - office and data utilities, legal and compliance helpers, privacy and security checks, and developer formatters. Every tool has its own page, and most can also be called from ChatGPT or Claude over MCP.