Skip to main content

Documentation Index

Fetch the complete documentation index at: https://agnt.mintlify.app/llms.txt

Use this file to discover all available pages before exploring further.

The Builder is a Claude agent that designs other Claude agents for you. It runs on agntdata’s own backend (not Anthropic’s managed runtime) so it can:
  • Maintain long-term semantic memory across sessions (pgvector-backed).
  • Keep an evolving workspace profile (WORKSPACE.md style) about your goals and conventions.
  • Read your existing APIs, integrations, and deployed agents.
  • Propose configuration changes that you approve before they go live.

​
Starting a session

Open Agents → Build with AI. The Builder remembers what you’ve worked on previously — start a new session per agent or topic.

​
Templates

When you start a session you can pick from a curated set of templates that pre-seed the conversation:
  • Customer success copilot — searches your CRM and answers product questions on Slack.
  • Outbound research agent — finds prospects via the LinkedIn data API and drafts personalized openers.
  • On-call triage agent — watches a Slack channel, classifies incoming requests, and escalates the urgent ones.
You can always start from scratch — templates are just shortcuts.

​
How proposals work

When the Builder wants to change a deployed agent’s configuration (system prompt, tools, model), it creates a pending proposal instead of editing the agent directly. You’ll see a side panel in the session UI with Approve / Reject buttons. Approving applies the diff to deployed_agents.config and (if the agent is already deployed) re-syncs the Anthropic registration so the next conversation picks up the changes immediately.

​
Memory

Two things are persisted:
  1. workspace_profile — a freeform markdown blob the Builder updates as it learns about your team, products, and preferences.
  2. memories — discrete embeddings of design decisions, user preferences, and rationale. The Builder retrieves the most relevant ones at the start of each turn so it doesn’t ask you the same question twice.
Both are scoped to your workspace; nothing leaks across workspaces.