Agents

The grid lists every agent in your account. Each card carries the agent name, model badge, last-updated timestamp, and a quick-action menu (Test, Edit, Duplicate, Delete). The top bar shows your plan limit and a + New Agent button when you have headroom.

• + New AgentOpens the four-step wizard — Template → Identity → LLM → Done. Hidden when you've hit your plan limit; the chip on the right shows current usage.
• FilterNarrows the grid by model, provider, or template. Filters stack — picking OpenAI + gpt-4o shows only agents that match both.
• SortReorders the grid by name, model, last updated, or last used. The default is most-recently-updated first.
• SearchLive filter against agent name and description. Combines with the filter chips above.
• Agent cardEach tile shows name, description, model + provider badge, last-updated time, and a menu with Test, Edit, Duplicate, Move to Folder, and Delete.

1. 1

   Pick a template

   Templates ship with a battle-tested system prompt and tone. Pick Website Chatbot to ground an agent on your site content, Order Status for an MCP-backed lookup, or Custom for a blank slate.
2. 2

   Set identity

   The name + description pre-fill from the template. Click Prompt Library to swap the system prompt for one of the curated presets. The description is what teammates see in the dashboard and what members see in the portal agent picker.
3. 3

   Connect a model

   Pick a provider, paste a key (or choose a saved one from Use saved key — see LLM Keys), set the temperature, and expand the Advanced panel if you need to tune max tokens, top-p, or output type.
4. 4

   Save and test

   The wizard finishes on a confirmation screen. Click Go to agent, then Test in the top-right to open the Playground with your new agent loaded.

• OverviewIdentity + model summary. Live status badge and the Agent ID chip you'll send as X-Agent-Id.
• PromptThe live system prompt. Edits are saved as a draft until you click Publish — see Playground for how draft vs published behave.
• KnowledgeQ&A pairs and short snippets the agent will always cite. Inline editor with a hit-count badge per entry once traffic flows.
• DocumentsUpload PDFs, Word, Markdown, or plain text. We chunk, embed, and retrieve the top-matching passages on every request.
• Web SourcesCrawl a public URL and use the resulting pages as knowledge. Useful for product help centres.
• MCP ServersConnect external tools and data sources via the Model Context Protocol. The agent discovers tools automatically; you control which ones are exposed.
• Token CostPer-million-token rates the platform uses to attribute spend to this agent. Keep these accurate so Analytics and Tracing match your provider bill.
• VersionsEvery publish creates a snapshot. Diff, roll back, or pin a member-portal agent to a specific version.

• Temperature0.0 – 2.0. Lower = more deterministic; higher = more creative. Templates ship with the right default for their use case.
• Max tokensHard ceiling for the response length. The default fits the agent's template; raise it for long-form assistants.
• Output typetext, markdown, json, or code. JSON output is validated server-side; invalid replies are auto-retried (up to 2) before erroring.
• StreamingOn by default for Pro+. Disable to receive a single response body — useful for batch jobs that don't need progressive UI.
• Max tool roundsCaps how many MCP tool round-trips the agent can take before the platform returns the final reply. Default 8 — enough for most multi-step lookups, low enough to bound cost.