Patterns

A reusable architecture catalogue of patterns and heuristics for agentic AI systems. These patterns are derived from evidence across our experiments and learnings.

exploring Feb 2026

Auto-Injection of History

When current session state is empty, automatically backfill chat context with recent stored history to preserve continuity.

Confidence: high
agentic-ai memory history
exploring Feb 2026

Context Injection

Collect relevant state from system stores and inject it into prompts as structured context to enable data-aware responses without fine-tuning.

Confidence: high
agentic-ai memory context
exploring Feb 2026

Dual-Path Routing

Classify user intent into a small set of routes and dispatch to specialized handlers to optimize responses for different request types.

Confidence: high
agentic-ai planning routing
exploring Feb 2026

Graceful Fallback

Wrap LLM calls with error handling that returns safe defaults so the system stays usable when generation fails.

Confidence: high
agentic-ai safety reliability
exploring Feb 2026

Heuristic-Driven Arbitration

Use a control point to adjust LLM-produced outputs based on deterministic safety signals from other components.

Confidence: high
agentic-ai control safety
exploring Feb 2026

Deterministic Guardrail

Use a non-LLM, rule-based agent to enforce safety constraints with predictable, explainable outcomes.

Confidence: high
agentic-ai safety guardrails
exploring Feb 2026

Implicit Tool State Mutation

Accumulate complex outputs by having a tool call mutate agent state during generation, then read the final result from that state.

Confidence: high
agentic-ai execution tools
exploring Feb 2026

Orchestrator-Managed Pipeline

Use a central orchestrator to run a fixed, dependency-ordered sequence of agents so analysis steps remain consistent and traceable.

Confidence: high
agentic-ai control orchestration
exploring Feb 2026

Prompt-as-Code Separation

Store prompts outside agent code and load them at runtime to improve maintainability and enable language variants.

Confidence: high
agentic-ai contract prompts
exploring Feb 2026

Structured Output via Schema

Define a formal schema for agent outputs and validate LLM responses against it to keep results machine-readable and reliable.

Confidence: high
agentic-ai contract structured-output
exploring Feb 2026

LLM Provider Abstraction Boundary

Introduce a provider-agnostic interface so agents can switch LLM backends without rewriting orchestration logic.

Confidence: high
patterns llm abstraction
exploring Feb 2026

Stabilise the Core Loop Before Adding Modalities

Defer multimodal features until the text-based agent loop and contracts are stable, to avoid compounding uncertainty.

Confidence: medium
patterns stability scope
exploring Feb 2026

Prompt–Schema Contracts as First-Class Architecture

Treat prompts, schemas, and structure locks as architectural contracts that constrain outputs and reduce drift.

Confidence: high
patterns prompts schemas
exploring Feb 2026

Derived Documentation from Evidence

Treat docs as derived state by regenerating AS_BUILT and DELTA_BACKLOG from executable evidence to prevent drift.

Confidence: high
patterns evidence documentation

Patterns are not just ideas—they are proven solutions to recurring problems in specific contexts.