Principles

Design principles that govern how agentic systems should be built. These are not suggestions -- they are constraints learned from production systems that fail when violated.

Agent Design

Principle	Core Idea
One Agent One Prompt One Purpose	Every agent has exactly one job
Specialization Compounds	Narrow focus multiplies capability over time
Architecture is Agent Performance	System design matters more than prompt tricks
A Priori Friendly Naming	Names should be guessable before reading docs

Context Management

Principle	Core Idea
Context Has a Sweet Spot	Too little context fails; too much drowns
Minimum Context	Give agents exactly what they need, nothing more
Avoid Context Pollution	Irrelevant context degrades agent performance
Documentation is Agent Memory	Written docs are how agents remember across sessions

Execution

Principle	Core Idea
The Plan is the Prompt	Well-structured plans execute themselves
Classification Precedes Planning	Know what kind of work it is before planning
State is Workflow Memory	Persistent state files carry context between steps
Prompts Are Evals	A prompt that works is also its own test

Delegation

Principle	Core Idea
Delegation Has a Sweet Spot	Over-delegation wastes tokens; under-delegation wastes capability
Isolation Enables Parallelism	Independent agents can run simultaneously
Domain Locking Scales Multi-Agent Safety	Confine agents to their domain to prevent conflicts

Quality

Principle	Core Idea
Trust But Verify	Agents produce; verification confirms
Your Work is Useless Unless Tested	Untested output is unshipped output
Cost Quality is a Knob	Trade off cost and quality deliberately, not accidentally

System Design

Principle	Core Idea
Fix the System Not the Issue	Patch the process, not just the symptom
Irreplaceable Work Lives at the Agentic Layer	Invest engineering time in the infrastructure, not the output
Prioritize Agentics	Agentic layer improvements compound across all work
Types Are Agent Contracts	Type systems enforce agent interface agreements