Agentic AI Design Patterns

Interactive simulations exploring resilience, routing, and orchestration patterns for LLM-powered systems. Click any pattern to try the live simulation.

Resilience Patterns
🔀
Multi-Region Circuit Breaker
3-layer resilience for LLM APIs: ADK retry, circuit breaker state machine, and cross-region failover. Watch circuits trip, cool down, and recover in real time.
Resilience Routing Interactive
Try simulation →
🧱
Bulkhead Isolation & Adaptive Concurrency
Per-tenant semaphore pools with AIMD token-rate control. Watch how one tenant's overload stays isolated while others serve normally.
Resilience LLM Infra In Progress
View repo →
🔗
Fallback Chain (Multi-Provider)
Cascading through Anthropic, OpenAI, and Gemini with provider-specific error classification. Visualize the waterfall, latency cost, and depth tracking.
Resilience Routing In Progress
View repo →
🔄
Request Coalescing / Singleflight
Deduplicates identical concurrent requests so millions of callers share one upstream LLM call. Watch requests collapse and fan-out on completion.
Resilience LLM Infra In Progress
View repo →
Safety Patterns
🛡️
Defense-in-Depth Safety Pipeline
Layered input/output filtering: Model Armor, Presidio PII redaction, IAM tool scoping, and schema validation. Watch requests flow through each guard stage.
Safety LLM Infra In Progress
View repo →
🔐
IAM-Enforced Tool Access with Budget
Infrastructure-level access control for agent tool calls. Least-privilege enforcement that doesn't rely on LLM judgment. Watch the allow/deny decision tree.
Safety Orchestration In Progress
View repo →
🎯
Red-Team Attack Classification
26+ adversarial attack types mapped to defense layers. Simulate prompt injection, PII exfil, jailbreak, and tool escalation — see which layer catches each.
Safety In Progress
View repo →
Memory Patterns
📜
Event-Sourced Session Persistence
Immutable event log with periodic checkpoints. Simulate crash, replay from checkpoint, and idempotent tool-call deduplication during resume.
Memory Resilience In Progress
View repo →
🏗️
Five-Tier Context Compaction
L1 (verbatim) through L5 (full transcript) memory hierarchy. Watch token budgets fill, compaction triggers fire, and tiers reshape over a 50-turn conversation.
Memory LLM Infra In Progress
View repo →
🧠
Cross-Session Memory Bank
Episodic, semantic, and procedural stores across sessions. Visualize memory consolidation, conflict resolution, and GDPR-compliant delete cascades.
Memory Governance In Progress
View repo →
Orchestration Patterns
📊
DAG-as-Manager / Agent Rails
Explicit state machine controls workflow; LLM operates within bounded nodes. Watch DAG execution with per-node budgets, step ceilings, and HITL escalation.
Orchestration Safety In Progress
View repo →
✋
Human-in-the-Loop Checkpoint
DAG execution pauses at defined checkpoints, waits for human approval, then resumes or aborts. Simulate approval queues and timeout handling.
Orchestration In Progress
View repo →
⏱️
Per-Node Tool Budget & Step Ceiling
Hard caps on tool calls per task with escalation on exhaustion. Compare "with rails" vs "without rails" showing runaway cost prevention.
Orchestration FinOps In Progress
View repo →
Determinism Patterns
🎛️
Six-Tactic Variance Stabilizer
Toggle semantic cache, structured output, grammar constraints, consensus voting, deterministic artifacts, and shorter output — watch variance metrics drop in real time.
Determinism LLM Infra In Progress
View repo →
💾
Semantic Cache with Embedding Lookup
Embedding-similarity lookup returns cached responses on near-match. Visualize 2D embedding space, similarity thresholds, cache hits/misses, and invalidation.
Determinism Resilience In Progress
View repo →
🗳️
N-of-N Consensus Voting
Run N parallel LLM calls, vote on the most common response. Visualize parallel fan-out, response variations, majority selection, and cost multiplier.
Determinism In Progress
View repo →
Governance & Observability
👁️
Eval-in-Production (Shadow + Golden Set)
Shadow-route N% of traffic to alternate models. Visualize comparison scoring, drift detection over time, and automatic rollback on threshold breach.
Observability LLM Infra In Progress
View repo →
🚀
Prompt Canary Rollout
Deploy new prompt versions 5% to 20% to 100% with automated regression checks. Visualize quality metrics, anomaly detection, and one-click rollback.
Observability In Progress
View repo →
💰
Cost Attribution & Anomaly Detection
Per-tenant/feature token cost tracking with 3-sigma spike detection and auto-throttle. Watch cost bars accumulate and anomalies trigger alerts.
FinOps LLM Infra In Progress
View repo →
Cross-Cutting
🏛️
7-Layer Architecture Stack
Complete enterprise AI platform decomposed into 7 layers. Click any layer to explore its patterns, inject failures, and see cascade effects across the stack.
Architecture LLM Infra In Progress
View repo →