📋Field ManualAI

Patterns

Reusable solutions to common problems in AI agent systems. Each pattern includes a problem statement, decision framework, implementation guidance, and real-world failure analysis. These are not theoretical — every pattern has been tested in production autonomous agent operations.

Core33KB original research · 5 strategies · 6 failure modes · 10 rules

Orchestration Patterns

5 task decomposition strategies for multi-agent systems: functional, data, perspective, stage, and recursive decomposition. Includes 6 failure patterns with recovery strategies and a 10-rule production playbook synthesized from Anthropic, OpenAI Swarm, Microsoft MagenticOne, and CrewAI.

Core26KB original research · Complete template · Decision tree

Task Specification

Templates and decision frameworks for writing agent task specs. Covers required sections (environment context, objectives, success criteria, constraints), the delegation decision tree, and idempotency patterns. Born from a real failure: a sub-agent that burned 78,000 tokens because of missing environment context.

Framework22KB original research · 7 levels · Decision matrix

Delegation Framework

A 7-level delegation spectrum from 'Tell' (execute exactly) to 'Delegate' (full autonomy). Includes a decision matrix mapping agent capability × task risk, goal-setting frameworks (SMART for agents), and practical templates for each delegation level.

ArchitectureProduction-tested · 3 layers · Schema included

Memory System Design

Three-layer memory architecture for persistent agent context: Layer 1 is a knowledge graph (entity-based facts with timestamps), Layer 2 is daily notes (timeline events), Layer 3 is tacit knowledge (patterns and preferences). Includes schema designs, retrieval strategies, and maintenance protocols.

Advanced13KB original research · 4 dependency types · 3 coordination models

Multi-Agent Coordination

Coordination models for multi-agent systems: centralized orchestrator, hierarchical delegation, and peer-to-peer patterns. Covers dependency types (FS, SS, FF, SF), handoff protocols, parallel execution strategies, and communication overhead management.

Meta28KB original research · CRISP framework · Evaluation loops

Agent Optimization

Meta-guide for improving agent performance: the CRISP prompt framework, constraint patterns, few-shot injection strategies, evaluation loops, meta-learning systems, and verticalization approaches. Covers prompt engineering, memory strategies, tool selection, and workflow chaining.

About Agent Patterns

What are the best AI agent orchestration frameworks?

The most widely-used AI agent orchestration frameworks in 2026 include Anthropic's Claude with tool use, OpenAI's Agents SDK (successor to Swarm), LangGraph, CrewAI, and Microsoft AutoGen. Each has trade-offs: Claude excels at complex reasoning tasks, the Agents SDK provides the simplest multi-agent patterns, LangGraph offers fine-grained control over agent state, CrewAI provides role-based agent coordination, and AutoGen supports diverse conversation patterns. The best choice depends on your coordination pattern — see our Orchestration Patterns guide.

How do you prevent AI agents from wasting tokens?

The primary causes of token waste are: missing environment context (agents rediscovering known constraints), unclear success criteria (agents going in circles), and over-delegation (spawning agents for tasks the orchestrator already loaded context for). Prevention strategies include: comprehensive task specs with environment blocks, explicit "done" conditions, delegation decision trees (delegate only when context transfer cost < execution cost), and token budgets per task. See our Task Specification pattern.