Agentic systems are AI workflows that decide which tools to call, when to loop, and when to hand off to a human. In financial services, they power underwriting assistants, claims triage, and document-heavy processes. This post outlines how current model and API releases support these use cases without overclaiming.
An agentic flow typically: (1) receives a user or system request, (2) chooses and calls tools (APIs, databases, calculators), (3) reasons over results, and (4) repeats or returns an answer. In regulated environments you also need audit trails, guardrails, and human-in-the-loop checkpoints.
Anthropic's Claude (including Claude 4 Sonnet) is strong at multi-step reasoning and tool use. In agentic evaluations it solved a significantly higher share of problems than earlier Claude models, and it supports structured tool-calling patterns that fit well with internal APIs and data sources. For financial use cases, that translates into assistants that can look up policy terms, run eligibility checks, or draft summaries step by step, with each step inspectable.
On the Databricks platform, Claude is available natively via AI Gateway — no external API setup required. ai_query('databricks-claude-sonnet-4', ...) invokes Claude directly in SQL or PySpark for batch inference workflows.
On the OpenAI side, function calling (and the newer structured outputs with strict: true for JSON Schema) lets you define tools and get model-generated arguments that you execute in your own code. That is the foundation for agentic behavior: the model proposes actions; your application runs them and enforces permissions and validation.
OpenAI's Agents SDK adds structure for building agents: tools (including MCP servers), tracing, and hosted tools like web search or file search. The Chat Completions API and Assistants API both support function calling, so you can implement agents with the interface that fits your stack.
Regardless of which model provider you choose, Databricks AI Gateway gives you a single endpoint to access all models — Claude, GPT-4o, Llama, Gemini — with unified governance, rate limiting, payload logging, and guardrails via Mosaic AI Gateway. Unity Catalog governs everything: data, models, functions, pipelines, vector search, and serving endpoints.
For data engineering, Lakeflow provides three products that work together:
A realistic pattern is an internal assistant that:
You keep the model in the loop for reasoning and drafting; your systems handle all data access and decisions that must be deterministic or compliant. That is the right division for production in financial services.
Regardless of provider, production agentic systems need: (1) clear tool boundaries and input validation, (2) logging of every tool call and result for audit, (3) human approval or review for high-stakes outputs, and (4) rate limits and timeouts so one run cannot spin forever. Mosaic AI Gateway provides guardrails (safety filtering, PII detection) and payload logging out of the box.
If you are exploring agentic systems for underwriting, claims, or operations, starting with a single workflow (e.g. "document in, summary + suggested actions out") and one provider keeps the pilot tractable. Once the pattern is stable, you can add more tools and steps or switch models via AI Gateway without code changes.