Real-world templates

The templates catalogue at /templates is the section of the platform we are quietly proudest of, because it is what closes the loop between reading about an agentic pattern and seeing it actually work against real models with real data. Every template in the catalogue is a production-grade pattern that we have either seen ship inside a real company or that mirrors a publicly documented architecture from a team like Uber QueryGPT, Snowflake Cortex, or Salesforce Agentforce. Nothing in the catalogue is a toy "hello world" — every entry could be the starting point of something a real team would ship to real users.

Anatomy of a template

Every template page is structured the same way, deliberately, so once you have used one you know how to use all of them. The page is split into four parts and each part exists for a specific learning purpose.

• The explainer is a one-page write-up of the problem the template solves, who in industry uses this pattern today (with citations and case-study links where possible), the architectural choices that went into it, and the trade-offs the author made. This is the bit you should read before doing anything else; it is also the bit most other agentic AI platforms do not bother shipping at all.
• The "Run as-is" button provisions the entire template — the agents, the knowledge bases pre-loaded with sample data, the prompt library entries, any required skills — into your own workspace in a single click. Once provisioned the template is fully yours; nothing is shared with the original, so you can break it freely.
• The guided tour is an overlay that walks you through every node of the swarm or every section of the agent configuration, explaining what each one does and why it is the way it is. The tour is keyboard-navigable and you can leave and resume it at any point.
• The "Fork and customise" path drops you straight into the Swarm Canvas or the Agent Builder with every piece editable and a small panel reminding you of what each node is supposed to do. This is where actual learning happens — the moment you change the system prompt of the planner and watch the downstream behaviour change.

What ships in the catalogue today

The template catalogue is curated rather than crowd-sourced, which means every entry has been reviewed by a human before being added, runs cleanly end-to-end, and ships with the explainer and guided tour above. The current catalogue covers the patterns we believe every serious practitioner of agentic AI should be able to build from memory, and it expands roughly every two weeks.

• RAG support bot demonstrates a full retrieval- augmented-generation pattern from end to end. The template ships with a knowledge-base ingestion pipeline pre-pointed at a sample documentation site, vector search wired into the agent's tool belt, citation rendering in the chat UI, and a fallback behaviour that tells the user when an answer is not grounded in the provided sources. Forking this template is the fastest path to understanding how production RAG actually fits together.
• SQL analyst (Text-to-SQL) mirrors the pattern made famous by Uber's QueryGPT and adapted by many enterprise analytics teams since. The agent generates SQL, the runtime parses it with a SQL AST validator, rejects anything that touches a table outside the configured allow-list, executes the query under the user's row-level security context, and returns both the rows and a natural-language summary. The template includes a sample database and a semantic-layer description so you can see how column meanings are conveyed to the model.
• RevOps swarm is a four-agent pipeline that researches a prospect, enriches the profile with public data, drafts a tailored outreach message, has a critic agent review the message against a configurable tone and length rubric, and finally sends it — gated by a human-in-the-loop approval. This template is the easiest way to internalise how multi-agent critic loops actually behave.
• Planner-Executor is the classic two-agent decomposition pattern: one agent receives the high-level goal and emits a structured plan; the other executes each step in sequence and reports back. The template includes a deliberately weak executor model so that you can watch the failure modes when the executor cannot follow the plan, then swap in a stronger model and observe the difference.
• Customer triage wires together a classifier agent that routes inbound messages into one of several categories, a per-category responder agent tuned to that category, and a human-in-the-loop approval node specifically on the refund branch. It is a small swarm that touches every major building block: routing, specialisation, and HITL.
• Code reviewer, web summariser, PII sanitiser, self-correction critic and several others fill out the single-agent end of the catalogue with focused examples of one pattern each. These are intentionally short so that you can read the entire agent configuration in a single sitting.

How to actually learn from a template

The single highest-leverage workflow on the platform is the loop of opening a template, reading the explainer, running it once without modification while the trace viewer is open in another tab, then forking it and changing exactly one thing — a prompt, a model, a tool — and running it again. Doing this two or three times in twenty minutes will teach you more about how agentic systems actually behave than two hours of reading. We recommend starting with the RAG support bot, then the planner-executor, then the RevOps swarm, in that order; the difficulty curve and the conceptual coverage line up well with that sequence.