Creating an agent used to mean picking tools from a list, connecting MCP servers, selecting playbooks, and configuring sandbox capabilities manually. It worked, but it was a lot of clicking for something that should be obvious from the description.
Now it's not.
What changed
When you describe what an agent should do, the generation step looks at your available MCP servers, playbooks, and secrets, and picks the ones that match. You write "research competitors using web search and save results to a spreadsheet" — the system selects the right search tools, attaches the right MCP server, and only requests the API keys actually needed.
No more connecting every MCP server "just in case" or manually cross-referencing which secrets go with which tools.
How it works
The generation prompt now receives your full resource inventory: MCP servers with their descriptions, playbooks with their intent definitions, and available secret names. The AI picks what's relevant and skips what isn't.
The key insight was required_secrets. MCP server templates list all
possible env keys at the provider level — an omnisearch server might
support six different API keys. But a specific task only needs one or
two. The AI now selects at the task level, not the provider level.
Progressive disclosure
We also reworked the agent creation form. Basic fields up front — name, description, task. Advanced configuration (model, timeout, max iterations, sandbox capabilities) tucked behind an expandable section.
Most agents don't need custom timeouts. The ones that do are one click away.
Try it
Go to Agents, hit "New agent", and describe what you want. Watch it wire everything up.