Create your first agent

Walk through the agent creation wizard end to end. The PIFSP-5 wizard is identity-first and mode-driven, so each step shows you a live preview of the agent being built.

The goal

An agent named "Hello", built through the wizard, with a working chat turn. By the end, you understand which fields matter at create time vs. which are best left at defaults.

Steps

1. Open the wizard.

   Sidebar -> Agents -> "New agent". The wizard opens at /agents/new.

2. Identity.

   • Name: human-readable. Hello.
   • Slug: URL-safe id, derived from the name. Override if you have a naming convention.
   • Description: optional; surfaces in the agents list and on the share page.

3. Mode.

   Pick the dispatch mode:

   • Direct (default): the agent calls tools directly within its turn. Best for most cases.
   • Sub-agent: the parent agent dispatches a sub-agent for tool work. Useful for very large tool sets.
   • Execute-tool only: the agent only sees find_tools plus execute_tools; pure discovery.

   Default direct.

4. Model and provider.

   Pick a model from the catalog. Only models whose provider has a linked key in the current environment appear. Without a linked provider, the picker is empty; jump to Add a provider key.

5. Prompt.

   Three options:

   • From scratch: write a system prompt in the text area.
   • From a template: pick from project-scoped prompt library entries (see Prompts).
   • Block-based: compose ordered prompt blocks for cache-friendly turns.

   For a first agent, start simple: You are a friendly assistant. Reply concisely. You can move to blocks once you understand the cache layers.

6. Tools (optional for first agent).

   Skip for now. The wizard auto-enables the standard meta-tool set (remember, recall, find_tools, execute_tools, spawn_bgo, etc.). Tools come in once you connect an entity or enable a skill.

7. Save and chat.

   Click "Create". The agent detail page opens; the first version snapshot is written. Click "Chat", type "Hi", watch tokens stream.

Verify

• The agent appears in the sidebar agents list.
• A first version exists (Versions tab shows v1).
• The chat turn streams tokens; the trace tab shows a span timeline; cost lands in the monitoring tab.

Test before publishing

Use Postman mode in the chat panel to inspect the assembled prompt and any tool calls. Save a test interaction as a Postman template (see Publish a Postman conversation template) so anyone on the team can replay.

Next steps

• Connect an entity (TypeScript) to give the agent tools.
• Version, canary, and roll back an agent for safe iteration.
• Run an A/B eval suite to compare versions on a fixed set.
• Set a per-agent budget cap before exposing the agent to traffic.