The 12 patterns
From Google Cloud’s agentic AI architecture guide:
| # | Pattern | Implementation in Genie |
|---|---|---|
| 1 | Single-agent | agents/<base agents> |
| 2 | Multi-agent sequential | pkg/workflow DAG |
| 3 | Multi-agent parallel | pkg/workflow fan-out |
| 4 | Multi-agent loop | pkg/reasoning Reflexion |
| 5 | Review and critique | pkg/safety plugin chain |
| 6 | Iterative refinement | pkg/reasoning Reflexion (same) |
| 7 | Coordinator | agents/financial_supervisor |
| 8 | Hierarchical task decomposition | agents/sme_loan_workflow |
| 9 | Swarm | (not implemented; intentional) |
| 10 | ReAct | pkg/reasoning ReAct |
| 11 | Human-in-the-loop | pkg/workflow saga + HITL |
| 12 | Custom logic | anything that doesn’t fit above |
Patterns I use most
Coordinator (pattern 7). A supervisor agent routes work to specialists. The dominant shape for finance and medical. In Genie: financial_supervisor calls analyzer, forecaster, anomaly_detector, etc. In Bodh: intake routes to questioner, test_planner, diagnostician.
HITL (pattern 11). Above the threshold, hand to a human. Payment over ₹2 lakh, KYC enhanced due diligence, claim disputes. The pattern is a saga with a manual checkpoint.
Review and critique (pattern 5). Generator agent + critic agent. The critic doesn’t generate; it scores. Used in Bodh’s reasoning_verifier after the diagnostician proposes a differential.
ReAct (pattern 10). Reason → act → observe → repeat. The default for “agent uses tools to figure something out.” Bounded by a max-iterations counter; without that bound, ReAct loops forever on hard problems.
The one I deliberately skip
Swarm (pattern 9). All-to-all communication between agents to converge on a solution. Useful in creative / brainstorming domains. Wrong for regulated finance — the audit trail becomes unreadable (which agent caused which output?). The cost is hard to bound (any agent can ask any other).
For regulated workloads, the audit clarity of coordinator + sequential is worth the constraint of not using swarm.
Patterns I’d add if I were re-writing the taxonomy
- Fallback (pattern 13): when the primary agent fails, route to a deterministic fallback. Critical for BCP; Google’s taxonomy doesn’t separate this from custom logic.
- Tier promotion (pattern 14): the dispatch decision based on the agent’s maturity (Sketch / Prototype / Beta / Production). Cross-cutting concern not just a pattern, but worth naming.
Where the taxonomy helps
When a new use case comes in, walking the 12 patterns in order is faster than starting from scratch. About 80% of customer use cases land in 1, 2, 7, or 11; the rest cluster around 4, 5, 8.
For the canonical Genie mapping, see docs/free-ai-mapping.md which threads the 12 patterns through the FREE-AI recommendations.
The taxonomy is a useful starting point. The implementations vary by domain; the conceptual model holds across them.