Launching Fall 2027
Agent Roles
Every system we install is run by a small set of standardized agent roles, each with a defined job, defined boundaries, and a clear handoff to the next. Here's how they work together.
The Orchestrator
GaiO
GaiO is the coordinating layer. It handles conversational intake, breaks work into the right tasks, sequences them across agents, tracks state as work moves, and assembles the final response. GaiO never bypasses the Guardrail Agent, never executes high-risk actions directly, and never expands its own permissions. GaiO is represented by the shield mark, not a character, since its job is coordination rather than a single visible task.
Bounded agents, not free-roaming AI
GaiO does not build agents that wander off and improvise. Every agent we deploy has a defined business purpose, defined inputs, defined outputs, defined tool permissions, defined review gates, defined escalation paths, defined logging, and defined rollback readiness. If an agent cannot be bounded that clearly, it does not go into a client's operation.
The standard pattern behind every agent
- Signal in
- Interpretation
- Policy check
- Action or review
- Log
- Monitor
- Improve
Four core families
Every agent we deploy belongs to one of four families. Some families cover a single focused role; others cover two agents that share the same operating posture.
Strategist
Coordinates the system and turns signal into recommendations.
Guardian
Watches the operation and enforces what is and isn't allowed.
Runner
Triages incoming work and routes it to the right place.
Builder
Prepares the work product and carries out approved actions.
The 7 agent roles
Each of the seven agent roles reports into one of the four core families below. Every role has an explicit list of what it is allowed to do, and what it is never allowed to do without a human in the loop.
Strategist
Orchestrator
Head Orchestrator
Conversational command layer and workflow coordinator.
Allowed
- Route tasks between agents
- Enforce workflow sequence
- Trigger human review
- Invoke approved sub-agents
- Choose an approved fallback path
Not allowed
- Bypass the Guardrail Agent
- Execute high-risk actions directly
- Self-expand its own permissions
Intelligence
Intelligence Agent
Analyzes patterns, scores outcomes, and recommends next steps.
Allowed
- Interpret data
- Compare options
- Score and recommend
- Produce insight for review
Not allowed
- Directly execute business-critical actions without separate approval
Guardian
Observer
Observer Agent
Monitors the operation, summarizes activity, detects issues, and reports out.
Allowed
- Read and summarize activity
- Recommend follow-up
- Flag issues
- Open an internal review item
Not allowed
- Contact customers directly
- Delete records
- Take autonomous external action
Guardrail
Guardrail Agent
Validates consequential actions before they happen, and blocks, pauses, or escalates when something doesn't check out. This role exists in every serious deployment we run.
Allowed
- Block a proposed action
- Pause a workflow
- Escalate to a human
- Require review
- Log and explain a policy issue
Not allowed
- Be skipped or overridden by any other agent
Runner
Router
Router Agent
Intake triage: classifies, tags, and routes incoming work to the right queue.
Allowed
- Classify incoming items
- Assign and tag
- Queue work
- Notify internal staff
Not allowed
- Give final approval on sensitive actions
- Make unrestricted edits to CRM records
Builder
Drafting
Drafting Agent
Prepares content, summaries, and structured outputs for review.
Allowed
- Draft and personalize content
- Propose next steps
- Prepare structured responses
Not allowed
- Send anything without authorization, unless explicitly pre-approved as low-risk
Execution
Execution Agent
Performs bounded, pre-approved actions inside a defined workflow.
Allowed
- Run narrowly defined workflow steps
- Trigger approved automations
- Update bounded internal records
Not allowed
- Take broad write access
- Grant itself policy exceptions
- Run unreviewed high-risk actions
Observer vs. Guardrail
This is one of the more important distinctions in how we design agent systems. Observer and Guardrail sound similar. They are not the same job.
Observer
Operational awareness. It watches the system and keeps everyone informed. It does not decide policy enforcement.
Guardrail
Policy enforcement and safety routing. It decides whether an action is allowed, needs approval, is too low-confidence, touches restricted data, or should be paused, blocked, or escalated.
Built to scale with your business
Small to medium enterprise
The seven agent roles operate directly inside your existing tools, with baseline governance already built in: defined permissions, mandatory logging, and human review gates on anything consequential.
Larger enterprise
The same agent roles operate under elevated governance: dedicated infrastructure options, stricter audit logging, multi-department routing, custom escalation paths, and deeper review-gate configuration for higher-stakes environments.
Want the full detail on how we govern these agents? See AI Security & Governance →
Book an Early Access Demo