What are the best agent platform setup options for a pro account?

Summary

• Pro accounts provide dedicated compute, multi-model routing, and advanced observability. The best setup aligns compute resources with your agent's latency and throughput requirements.
• Start with serverless deployment for flexible latency tolerance. Use provisioned compute for agents that need consistent response times or continuous background work.
• Set up multi-model routing to balance cost and performance: use frontier models for complex reasoning, fine-tuned or open-source models for simpler tasks.
• Integrate your required tools (MCPs for GitHub, Jira, Slack, Google Drive) and Skills at setup time, not after deployment. Test them against realistic data volumes and error conditions.
• Use Unity Catalog and Service Policies to enforce fine-grained access control per agent. Set budget limits and cost alerts before deploying to production.
• Enable observability from the start: use Unified Trace Table for debugging, monitor error rates, and implement alerting on SLA violations.

What are the best agent platform setup options for a pro account?
Pro accounts provide the full suite of Agent Bricks capabilities. Setup focuses on three decisions: what compute infrastructure to use, which models to route to, and what tools and integrations your agents need. The best setup depends on your response time targets, cost constraints, and operational complexity.

Compute infrastructure options

Pro accounts support both serverless and provisioned compute. Serverless is best when response time tolerance is flexible and you want to avoid managing infrastructure. Provisioned clusters are better for agents that need consistent, predictable response times or run continuous background workloads.
For most agents, start with serverless and measure actual response times. If performance doesn't meet your SLA or your agents run constantly, switch to provisioned. You can use both: serverless for exploratory agents and light workloads, provisioned for production-critical agents that serve many concurrent users.
Size provisioned clusters based on your agent's model size and expected concurrency. A complex reasoning agent needs different resources than a lightweight classification agent.

Model and provider setup

One advantage of pro accounts is access to multiple models and providers. Instead of betting on one frontier model, set up routing that uses different models for different tasks. Your agent might call Claude for complex reasoning, GPT for code generation, and a fine-tuned model for classification.
Register models through Unity Gateway so you get centralized cost tracking, rate limiting, and access control. This prevents one agent from monopolizing budget or hitting provider rate limits. Store provider credentials (API keys) in Unity Gateway, not in agent code.
Evaluate latency vs. quality for each model. Frontier models are slower but higher quality. Smaller models are faster and cheaper. Test your agent with each model on representative data before production.

Tool and MCP integration

Agents are only useful if they can interact with your systems. The best setup integrates your required tools at deployment time, not reactively when an agent fails. Common integrations for pro accounts include GitHub (for code changes), Jira (for task management), Slack (for notifications), Confluence (for documentation), and Google Drive.
Each tool integration (via MCP or Skill) requires setup: authentication, permission configuration, rate limiting, and testing. A pro setup has these configured before the agent goes live. Test tool calls with realistic data volumes and error conditions. An agent that calls tools repeatedly can generate surprising API costs or hit rate limits.
Use Service Policies to control which agents can call which tools. A coding agent might have GitHub access, but not Jira. A financial reporting agent might have database access but not external integrations.

Governance and cost controls

Pro accounts need centralized governance. Set up Unity Catalog with role-based access control so agents inherit user permissions (row-level and column-level). Define which teams own which agents. Use audit logging to track all agent activity.
Set cost budgets and alerts before deploying. Track spending per team, per agent, and per user. Implement hard spend caps for agents that might have runaway queries. A single agent that queries large data volumes without proper filtering can generate unexpected bills.

Observability and monitoring

Pro setups require observability infrastructure. Enable the Unified Trace Table to capture every model call, tool invocation, and agent decision. Set up dashboards to track agent latency, error rates, tool success rates, and cost.
Implement alerting on SLA violations: if an agent's latency breaches its target threshold, or error rate becomes unacceptable, or daily cost exceeds budget, alert the team. These alerts should go to Slack or another notification channel your team monitors.
Use observability data to iterate. If an agent is slow, is it the model, the tool, the routing logic, or the data query? Trace data answers these questions.

Configuration checklist

Item	Decision	Notes
Compute type	Serverless or provisioned?	Start with serverless unless very low latency is required
Cluster size	Cores and memory	Base on model size and concurrency
Model routing	Which models for which tasks?	Register through Unity Gateway, not hardcoded
Tool integrations	GitHub, Jira, Slack, custom APIs?	Test before production
Access control	Who can use this agent?	Use Service Policies
Cost budget	Per team, per agent, daily/monthly?	Set alerts and hard caps
Observability	Trace logging enabled? Dashboards configured?	Mandatory for production
Error handling	Graceful degradation or retry logic?	Test before deployment

FAQs

Should pro agents always use frontier models?

No. Frontier models are slower and more expensive. Use them for tasks that require reasoning or complex understanding. Use smaller models or fine-tuned models for classification, summarization, or other high-volume, low-complexity tasks. Multi-model routing is the pro pattern.

How do I optimize agent cost?

Track cost per task through Unity Gateway, not just raw token count. Some tasks benefit from cheaper models without sacrificing quality. Use observability data to identify expensive operations and opportunities to use smaller models or batch requests.

How do I know if my agent is performing well?

Define SLAs before deployment: response time targets, error rate tolerance, and tool success rate. Use Unified Trace Table data to measure against these SLAs weekly. If metrics degrade, investigate through observability data.