Procedure
The step-by-step instructions the agent follows. Plain language that operations and compliance can both read.
Agent Studio
Agent Design
Four ways to build an agent, one profile at the end. Whether you start from a template, answer a wizard, draw on a canvas or edit YAML directly, the result is an Agentic Orchestration Profile: a versioned file the platform can govern.
Build paths
Four ways in, one format out.
The AOP is the single source of truth for what the agent may do. Every build path produces the same file, so your compliance team reviews one format regardless of how it was made.
Template start
Pick a domain template and customise it. The fastest path when a template covers your job.
No-code wizard
Answer plain questions about what the agent does, what tools it needs and what it must never do. The wizard drafts the profile for review.
Flow canvas
Draw the conversation flow when the job has branches, handovers or conditional logic. The canvas saves as the same standard profile.
Direct YAML edit
Write or paste the AOP directly for full control. Validation runs on every save.
The profile
What goes into an Agentic Orchestration Profile.
An AOP is a versioned YAML file checked into git. It lists what the agent does, what tools it may call, what checks it must pass and how much freedom it has.
Tools
The tools the agent may call: CRM lookups, payment writes, knowledge base queries. Every tool resolves against the connector hub.
Guardrails
Checks that fire after every generation: PII detection, contact rules, compliance boundaries. Universal guardrails are added by the platform.
Autonomy
Shadow, assisted or autonomous. The agent starts restricted and earns its way up through evals and promotion.
Model strategy
Which model to use and where to route it. The model router applies tenant policy on top.
Version control
Every save creates a new version. Promotion requires a signed tag. You can diff any two versions and roll back.
Tool wiring
Connect the agent to your systems.
The Studio validates every tool against the connector hub. A tool that does not resolve blocks the profile from passing readiness checks.
MCP tool import
Import tools from an MCP server. Imported tools start gated as writes until you mark them read only.
Default deny
A tool not in the profile is refused at runtime. A tool not marked read only is gated as a write and sent to the approval queue.
Connector hub
All tools dispatch through one hub. An in-process connector and a remote MCP server are interchangeable: a registration change, not a code change. The hub enforces tool-kind classification so a read tool cannot be used as a write.
Guardrail configuration
Choose what the agent must check on every step.
Templates come with domain guardrails pre-wired. You can add, remove or tune them. Universal guardrails (PII, safety, audit) are always present and cannot be removed.
Input guardrails
Check the user input before the model sees it. PII detection, prompt injection defence and content policy.
Output guardrails
Check the model output before the agent acts on it. Tone, accuracy, regulatory language and citation requirements.
Tool guardrails
Check tool inputs and outputs. Contact window enforcement, payment limits and data integrity validation.
Who designs
Operations builds. Engineering reviews.
Why it matters
The people who know the job best design the agent. Engineers review the tool connections and the guardrail coverage. Neither group works alone, and the platform records who did what.
Design an agent in one sitting.
Bring a workflow and we will walk through the wizard together, from blank page to a profile ready for testing.