# Custom agents > Attio's Custom agent handles the work that rules can't: the tasks that need a judgment call. It runs the task you describe. You write the instructions, choose the model, choose which tools it can reach, and define the shape of what comes back. It runs as a block in any workflow, or as an AI attribute on a record. It replaced the prompt completion block. ## What is a Custom agent? A Custom agent is a configurable AI agent inside Attio. Unlike a fixed-purpose agent such as Web Agent, you define what it does: - **Instructions** — what the job is, in plain language. A separate system prompt defines the agent's role, rules, and behavior. - **Context** — the agent has no data by default. You pass context in using variables from earlier blocks, and by enabling Attio read tools. - **Model** — choose between OpenAI (GPT), Google (Gemini Flash and Pro), and Anthropic (Claude Sonnet and Opus). - **Tools** — Attio read-only tools (records, lists, notes, emails, call recordings, meetings, tasks, workspace members), an optional web-access toggle, and any MCP-connected app with per-tool selection. - **Run as** — the workspace member or workflow whose credentials and permissions the agent uses. - **Output** — plaintext, or a structured JSON schema where each property becomes its own output variable for later blocks. ## Where it runs | Surface | What it does | | --- | --- | | Workflow block | Runs your instructions as a step, with its output available to every block after it | | AI attribute | Autofills an attribute from your own prompt (formerly the prompt completion type) | ## Why structured output matters Because each schema property becomes a separate workflow variable, a Custom agent is a composable building block rather than a chat response. Later blocks can branch, score, route, or draft on individual fields it returns. ## Example agents teams build - **Lead scorer** — scores inbound against your own ICP criteria - **Meeting prep** — reads notes, emails, and calls, then writes a brief - **Deal risk** — flags deals going quiet and explains why - **Call coach** — reviews a call against your own scorecard - **Support triage** — reads a ticket, tags it, routes it - **Competitor watch** — checks a site and pricing, notes what changed - **Onboarding plan** — turns a kickoff call into the next set of tasks - **Win/loss notes** — extracts the reason in your own categories - **Overnight digest** — scans the book and surfaces what needs attention - **Expansion signal** — spots usage growth and drafts the upsell note ## Trust and governance - **Agents act as identities.** A Custom agent runs as the member you assign, with exactly their permissions. - **Every run is on the record.** Open any run to see what the agent read, decided, and did. - **You keep control.** Pause, cancel, resume, or replay any run. ## Open and extensible - Every prompt, block, and agent is readable and editable — nothing is a fixed black box - Runs over Universal Context, the deepest context layer in agentic CRM - Reaches Notion, Linear, Fin, and any other tool over MCP - Extend with custom JavaScript, the Attio App SDK, and your own blocks ## Common Questions **How is a Custom agent different from Web Agent?** Web Agent has one job: research the open web. A Custom agent does whatever you instruct, with the model and tools you select. **Does it see my data automatically?** No. The agent has no data by default. You pass context in with variables, or enable specific Attio read tools. **Which models can I use?** OpenAI GPT models, Google Gemini Flash and Pro, and Anthropic Claude Sonnet and Opus. **Can it use tools outside Attio?** Yes. Turn on web access, and connect any MCP app, choosing which of its tools the agent may call. **What happened to prompt completion?** Custom agents replaced it, adding model selection, tool selection, and structured output. ## Getting Started - Start for free - Custom agents product page - Workflows block library ## Related pages - Web Agent - Workflows and revenue agents - Ask Attio - Complete platform overview