Guides

Already have agents running on Letta? Export them to YAML, check them into git, and manage them with lettactl going forward. No need to rewrite configs from scratch.

Source agent prompts, memory blocks, tools, and folder files from Supabase Storage buckets. lettactl downloads content at deploy time so your agent configs stay declarative while your content lives in the cloud.

Let users upload files from your frontend to a Supabase bucket, then have your agents pull from that same bucket. A shared bridge between your UI and your agent fleet.

Deploy agent changes to isolated canary copies first. Test with real traffic, then promote to production or roll back — without touching your live fleet.

Manage thousands of agents across tenants using tag-based filtering. B2B and B2B2C patterns for SaaS platforms built on Letta.

Let your agents analyze their own memory health. Detect stale data, overcrowded blocks, redundancy, and missing knowledge — using the agents themselves as the diagnostic tool.

Use first_message to prime agents with initial context on creation. When agents drift or need retraining, delete and redeploy them, or use bulk messaging to recalibrate your entire fleet in place.

Send the same message to hundreds of agents at once using glob patterns, tags, or config files. Bulk messaging runs 5 agents concurrently with real-time status output — useful for fleet-wide recalibration, announcements, and multi-tenant operations.

When you embed Letta agents in a web application, you cannot give them direct write access to your database, storage, or external services. A prompt injection, a hallucinated parameter, or a misunderstood instruction can mutate data, delete records, or bypass authorization — and there's no undo. This guide covers the mediation pattern: agents get read-only tools, and all writes go through your app's API layer where you control validation, auth, and audit.

Connect your lettactl-managed agents to multi-channel messaging platforms like Telegram, Slack, and Discord using LettaBot. Define everything in one fleet YAML and deploy with a single command.

Use conversations to give each agent multiple isolated message histories — perfect for per-user threads, A/B testing prompts, or separating hook and b-roll interactions. Conversations share the agent's archival memory, blocks, and tools while keeping messages completely separate.

Monitor async agent runs in real-time. Use watch mode for broad fleet monitoring, or the track command for focused parallel run tracking that auto-exits on completion — perfect for CI/CD pipelines and scripted workflows.