Team Architecture Patterns

A single hardened agent is a specialist. A team of hardened agents is a capability. The difference between the two is architecture — and architecture is where most deployments fail.

The AT track teaches three orchestration patterns: hub-and-spoke (lead dispatches to specialists), pipeline (output flows sequentially through stages), and debate (agents argue toward consensus). Each pattern has a cost profile, a failure mode, and a sweet spot. Hub-and-spoke costs 3-5x a single agent but handles 80% of multi-agent tasks. Pipeline costs 2-3x but only works for sequential workflows. Debate costs 5-10x but catches nuances that single-pass review misses. The mistake is choosing a pattern because it sounds sophisticated. Choose based on your task dependency graph.

1. Map Your Task Dependency Graph Before selecting a pattern, enumerate every subtask and draw the dependencies. Tasks with no dependencies run in parallel (hub-and-spoke). Tasks with strict ordering run in sequence (pipeline). Tasks that require mutual evaluation run in debate. Most real workflows are hybrids — parallel research feeding into sequential synthesis feeding into a debate review. Map it, then select the pattern for each phase.
2. Define Roles with Explicit Boundaries Every agent in a team gets a role definition structured like a job description: title, responsibilities, deliverable format, and what the agent must NOT do. "Research Agent — analyzes source materials, outputs structured JSON brief, does not write final copy." Boundaries prevent scope creep, which is the primary source of inter-agent conflict and token waste.
3. Implement Quality Loops AT quality loops add a dedicated critic agent that reviews specialist output before the lead aggregates it. Research shows this improves accuracy from approximately 45% to 55% across three critique rounds. The critic evaluates on defined dimensions — faithfulness, completeness, format compliance — and either approves or returns with specific revision instructions. Three rounds is the sweet spot. Diminishing returns after that.