Agents

Purpose: Specialize a model with a system prompt, optional tools, and optional RAG.

Key fields

1. Agent Name / Description — used for lookup in lists and in the API (model: "<agent-name>").
2. Agent Prompt — your system prompt (persona, rules, success criteria).
3. Agent Model — pick the named model you saved on the Models screen.
4. Use Tools (MCP) — enable and then pick specific servers and functions the agent may call. Principle of least privilege.
5. Use Embeddings (RAG) — toggle and select one or more collections to search and inject into the agent’s context.
6. Publish Agent (coming soon) — exposes to your account’s shared lists (and, if enabled elsewhere, to Slack). Not public outside your account and shared list.

Tool selection

When you enable Use Tools (MCP), you can select specific MCP Servers and functions the agent may call. In the dropdown, each item corresponds to an MCP Server. If no specific functions are selected, all functions will be provided to the model to use.

You may optionally refine the tools that are provided to the model. Some servers have many tools, and you may not want to provide all of them to the model. Too many available tools will waste tokens and could cause the model to make choices that may not be aligned with your intentions. Click "Select Functions" to choose the functions you want to make available to your agent.

Embedding Collections (RAG)

After you upload documents to the platform, you can select which collections you want to make available to your agent.

Advanced Options

• Combine System Prompts

  When enabled, the agent's system prompt will be combined with any system prompts passed in from external tools. The agent's prompt comes first.

• Convert System Prompts to User Messages

  When enabled, any additional system prompts from external tools will be converted to user messages instead of being treated as system prompts.

• Advanced Parameters

  When set, you can override the model's default parameters for this agent.

Sharing agents

Publish an agent and share it with another user (by username or email). Sharing has two modes:

• Read‑only — the recipient can run the agent and (depending on configuration) see it in their shared list.
• Editable — granted by sharing with edit access. The recipient can change a curated subset of the agent's configuration (description, system prompt, model options, tools, RAG collections, and voice), but not the underlying model, name, or response‑processor agent. Edits write back to your copy of the agent, and any tools/collections/voices the recipient picks must be ones you own — they choose from your resources, surfaced through dedicated owner‑resource listings.

Recipients can also run a shared agent directly from chat completions by targeting model: "shared/<ownerUsername>/<slug>".

Programmatic tool calling

Programmatic tool calling (Claude code‑execution tool use) is configured on the model, not the agent — an agent inherits it from its model. It is enabled automatically for Claude Sonnet/Opus 4.6+. See Models → Programmatic tool calling.

Design guidance

• Keep prompts explicit about sources of truth and cite/tool behavior (e.g., “always ground answers in RAG or tool output”).
• Create separate agents per job: e.g., web-news-finance vs code-reviewer.
• Parameter overrides (if supported) live here when you need behavior different from the base model config.

API parity

• List: GET /v1/core/agents
• Add: POST /v1/core/add/agent · Update: PUT /v1/core/update/agent
• Delete: POST /v1/core/delete/agent
• Share (editable): POST /v1/core/agents/:username/:slug/shares with editable: true
• Run via OpenAI‑compatible chat: POST /v1/chat/completions with model: "<agent-name>"
  See Agents and Chat Completions