pattern library · v1 · 23 patterns · 61 curated case studies
Agentic System Patterns
Evidence-backed building blocks for production AI systems, distilled from 1,780 real-world case studies. Start with the field report; go deep through the patterns; verify anything against the case studies behind it.
The Field ReportStandardized Bricks, Custom Buildings — what the whole corpus taught us: the thesis, the load-bearing blocks, the recurring decisions, what breaks, and how to think about boundaries. Read this first.Pattern library
Each page: definition, applicability signals, architecture variants, tradeoffs, gotchas with resolutions, when not to use it, an operational checklist, and real implementations. See also the vocabulary and the boundary matrix (which pattern applies when two overlap).
Orchestration
How work is divided, routed, and coordinated
tool-use-orchestration64.9% · 643 systemsTool-use orchestration is the control loop that lets a model select a bounded external capability, supply validated arguments, execute it through software, and return the result to the model or workflow.multi-agent-orchestration33.5% · 332 systemsMulti-agent orchestration decomposes a system into specialized model-driven roles and coordinates their routing, delegation, state, and returned results.prompt-chaining5.2% · 52 systemsPrompt chaining composes a known sequence of model calls where one stage's result becomes structured or contextual input to the next.durable-execution9.4% · 93 systemsDurable execution preserves workflow progress across process failure, retries, long waits, and asynchronous events.
Quality & Evaluation
How teams know the system works
eval-driven-development47.9% · 475 systemsEval-driven development makes representative evaluation data, explicit criteria, and release evidence part of the product-development loop.llm-as-judge12.1% · 120 systemsLLM as judge uses a model to apply an explicit rubric or checklist to another model's behavior, conversation, or artifact.observability-tracing33.2% · 329 systemsObservability and tracing records an agent's model calls, tool use, retrieval, workflow transitions, and outcomes as a connected execution history that operators can inspect at both single-run and aggregate levels.data-quality-loop9.9% · 98 systemsA data quality loop turns production errors, human feedback, audits, and changing source data into corrections to the knowledge, examples, labels, schemas, or evaluation sets used by an AI system.prompt-versioning6.0% · 59 systemsPrompt versioning treats system instructions, task prompts, examples, and related policy text as named release artifacts with ownership, history, deployment controls, and evaluation links.
Knowledge & Context
What the model sees and remembers
rag-retrieval30.7% · 304 systemsRAG retrieval selects source material at run time and places the relevant evidence into a model's working context.agent-memory18.5% · 183 systemsAgent memory persists and retrieves task-relevant state beyond the immediate model call.context-engineering8.4% · 83 systemsContext engineering is the discipline of selecting, structuring, refreshing, and isolating the information available to a model at each decision.context-compaction9.9% · 98 systemsContext compaction replaces or removes accumulated interaction state while preserving the information needed for subsequent decisions.
Governance & Safety
Who may do what, with which data
human-in-the-loop-gate40.0% · 396 systemsA human-in-the-loop gate is an explicit control boundary where automation pauses, escalates, requests approval, or routes a decision to a responsible person.safety-policy-enforcement21.3% · 211 systemsSafety-policy enforcement places deterministic authorization, compliance, and business-rule controls around probabilistic agent behavior.guardrails-io-filtering10.0% · 99 systemsGuardrails I/O filtering validates, blocks, transforms, or annotates content as it crosses a model or agent boundary.privacy-preserving-processing8.9% · 88 systemsPrivacy-preserving processing limits exposure, retention, movement, or reuse of sensitive data while still enabling an AI workflow.structured-output-contract15.9% · 158 systemsA structured output contract requires a model-produced result to conform to a machine-readable shape that downstream software can validate and consume.
Models & Economics
Which model runs, at what cost and speed
model-routing17.1% · 169 systemsModel routing selects a model or model configuration for a request, step, or workload according to capability and operating constraints.cost-aware-routing11.3% · 112 systemsCost-aware routing assigns model, inference, or workflow paths according to expected cost subject to a required quality and service envelope.latency-optimization10.7% · 106 systemsLatency optimization reduces elapsed or perceived response time across an agent's full critical path.fine-tuning-adaptation9.1% · 90 systemsFine-tuning adaptation changes model parameters using task, domain, style, or preference data so recurrent behavior is learned rather than supplied entirely at inference time.fallback-degradation6.1% · 60 systemsFallback degradation preserves a bounded useful behavior when the preferred model, tool, agent, or workflow path fails.
Case studies
The 61 systems the pattern pages cite as representative implementations — each with problem, architecture (with diagram), data flow, alternatives considered, tradeoffs, and gotchas, every claim quoted from the source. Browse or search the full corpus of 1780 case studies.
Corpus analysis
The quantitative layer behind everything: adoption, convergence, problem taxonomy, and the ranked index of what goes wrong in production.