The production agent stack is the decision.

Models

Provides the reasoning, language, vision, and generation capabilities behind each agent.

Example: A service agent understands a customer request and drafts an accurate response.

Routing

Balances quality, speed, availability, and cost for every request.

Example: A routine order lookup uses a fast model while a complex dispute is sent to a frontier model.

Orchestration

Turns a request into a dependable sequence of steps, decisions, and actions.

Example: A claims agent gathers documents, validates coverage, calculates a recommendation, and requests approval.

Agent-to-agent

Lets specialist agents collaborate on work that spans teams or systems.

Example: A sales agent asks pricing, inventory, and legal agents to assemble a compliant proposal.

Connectors

Gives agents governed access to the systems where customer and business data lives.

Example: An account agent reads Salesforce history and checks an SAP order without manual re-entry.

Grounding

Keeps answers anchored in current, approved enterprise information.

Example: A policy assistant cites the latest internal policy instead of relying on model memory.

Tools & APIs

Allows agents to complete work, not just recommend what a person should do next.

Example: A support agent issues a refund through an approved commerce API after validating eligibility.

Sessions

Preserves the state of an interaction across multiple steps and channels.

Example: A customer begins a return in chat and completes it later by phone without starting over.

Memory

Enables useful continuity while retaining control over what an agent remembers.

Example: A relationship agent remembers communication preferences and approved portfolio objectives.

Runtime

Runs multi-step agent work reliably beyond a single model response.

Example: A procurement workflow continues processing while waiting for an approval or supplier response.

Scale & isolation

Maintains performance during demand spikes while separating users and workloads.

Example: A retail assistant handles holiday traffic without one tenant affecting another.

Sandbox

Lets agents run code or process files without exposing production environments.

Example: A finance agent analyzes an uploaded spreadsheet inside an isolated execution environment.

Release controls

Enables controlled launches, versioning, testing, and rollback of agent changes.

Example: A new prompt is released to 10% of users and rolled back automatically if quality drops.

Identity

Establishes who is asking, which agent is acting, and who owns each action.

Example: An employee agent signs in through corporate SSO and every action is attributed to that identity.

Authorization

Restricts agent access to the permissions granted to the user and the task.

Example: A manager can approve an expense while the agent cannot access unrelated payroll records.

Policy & safety

Enforces business, safety, privacy, and regulatory rules during execution.

Example: A healthcare agent blocks unsupported clinical advice and routes the case to a clinician.

Human escalation

Keeps people in control of high-impact, uncertain, or exceptional decisions.

Example: A payment above a threshold pauses until an authorized employee approves it.

Audit

Creates evidence of what the agent knew, decided, and did.

Example: An auditor can reconstruct the data, policy, model version, and approval behind a decision.

Evals

Measures whether agent quality and safety remain acceptable as the system changes.

Example: A release is blocked when its refund-policy accuracy falls below the approved threshold.

Observability

Shows where an agent is slow, expensive, inaccurate, or failing.

Example: Operations traces a delayed answer to a slow CRM connector rather than the model.

Incident response

Limits customer impact and restores a safe service when something goes wrong.

Example: A faulty action tool is disabled globally and affected transactions are identified for review.

Registry

Makes every agent discoverable with clear ownership, purpose, status, and dependencies.

Example: Security can identify all agents using customer data and contact the accountable owner.

Cost controls

Keeps agent spending predictable and attributable to teams, customers, or use cases.

Example: A business unit receives a budget alert before an unexpected traffic spike exceeds its quota.