Pattern Vocabulary

Seed vocabulary for Stage 3 extraction. Reuse these names before proposing new ones.

Pattern name	Status	Notes
multi-agent-orchestration	accepted	Coordinating multiple agent roles or workers.
model-routing	accepted	Selecting models by task, cost, latency, or quality.
human-in-the-loop-gate	accepted	Explicit human approval, review, or escalation gates.
rag-retrieval	accepted	Retrieval-augmented generation over external knowledge.
guardrails-io-filtering	accepted	Input/output constraints, filters, or policy checks.
agent-memory	accepted	Persistent or session memory for agent behavior.
durable-execution	accepted	Durable workflow execution, checkpointing, replay, or resumable orchestration.
eval-driven-development	accepted	Evaluation harnesses guiding development or release.
prompt-versioning	accepted	Versioned prompt artifacts and change control.
fallback-degradation	accepted	Graceful fallback under failures or uncertainty.
semantic-caching	accepted	Reusing semantically equivalent model results.
context-compaction	accepted	Summarizing or compressing long-running context.
tool-use-orchestration	accepted	Controlled use of external APIs, functions, or tools.
retrieval-ranking	accepted	Ranking retrieved items before generation or action.
hybrid-search	accepted	Combining lexical and vector search.
query-rewriting	accepted	Rewriting user input for retrieval or tool execution.
structured-output-contract	accepted	Schema-constrained model outputs.
llm-as-judge	accepted	Model-based grading or comparison.
confidence-thresholding	accepted	Routing by confidence scores or uncertainty.
synthetic-data-generation	accepted	Creating generated examples for training or evals.
model-distillation	accepted	Transferring behavior from larger to smaller models.
fine-tuning-adaptation	accepted	Customizing a model with domain data.
prompt-chaining	accepted	Multi-step prompt workflows.
observability-tracing	accepted	Logging, tracing, or monitoring LLM workflows.
cost-aware-routing	accepted	Cost controls through routing, caching, or batching.
latency-optimization	accepted	Reducing response or inference latency.
safety-policy-enforcement	accepted	Enforcing safety, compliance, or business policy.
data-quality-loop	accepted	Feedback loops for improving datasets.
canary-rollout	accepted	Gradual deployment or release comparison.
batch-offline-processing	accepted	Offline or async LLM processing patterns.
privacy-preserving-processing	accepted	Redaction, isolation, or privacy-aware handling.
async-message-orchestration	accepted	Agent work decoupled via durable messages/queues/events (SQS, Kafka, EventBridge, semantic event buses) instead of direct calls. Accepted 2026-07-11: 27 projects, reviewer semantic review passed.
workflow-decomposition	accepted	Splitting a workflow into smaller steps or sub-agent responsibilities with bounded contexts. Accepted 2026-07-11: 16 projects; usage settled on the literal meaning (the original narrow-slice-scoping usage is the flagged outlier, re-mapped in analysis normalization).
context-engineering	accepted	Curating what enters model context: instruction files, context budgets, pruning, sub-agent context isolation. Broader than context-compaction. Accepted 2026-07-10 (9 supporting projects, reviewer adjudication in forward-batch6-pass).
forward-deployed-engineering	accepted	Embedded engineers own customer outcomes and translate enterprise workflows into production AI deployments. Accepted 2026-07-11: 3 projects (threshold), category delivery/operations.