CONCEPT Cited by 1 source

Agent fleet architecture

Definition

A system design where a fleet of specialized agents — each tuned to a narrow domain or input type — replaces a single generalist agent. A router directs incoming work items to the appropriate specialist based on metadata (source type, category, tenant). Each agent carries its own context (prompts, tools, filtering rules, behavioral baselines) that would be impossible to fit into a single-agent context window.

Why specialize?

A single agent reasoning across heterogeneous inputs lacks the domain-specific context to make accurate decisions. At Databricks, a generalist triage model produced a 50% escalation rate — essentially random noise. Splitting into 17 source-specific agents (each knowing its source's false-positive patterns, behavioral baselines, and enrichment needs) reduced escalation to 3.2% with ~10× higher true-positive rates (Source: sources/2026-07-06-databricks-scaling-security-alert-triage).

Design principles

1. One agent per input domain — the context needed for accurate reasoning is domain-specific.
2. Shared utilities layer — common framework for invocation, retries, cost tracking, tracing, and result persistence.
3. Shared service agents — cross-cutting capabilities (e.g., threat intelligence lookup) exposed as callable sub-agents.
4. Deterministic pre-filtering — rules handle predictable cases before the LLM sees anything.
5. Unified evaluation — all agents measured against the same ground-truth methodology.

Seen in

• sources/2026-07-06-databricks-scaling-security-alert-triage — 17 source-specific security triage agents at Databricks.