Back to archive

Guide to the Agent Design Paradigms Series

A guide to stable Agent system design paradigms: context, memory, tools, permissions, collaboration, compaction, recovery, and auditable execution.

Guide to the Agent Design Patterns Series

This series is for ideas that last longer than any single concrete Agent project.

Build Harness is better suited to step-by-step engineering construction: Provider, Loop, Tool Runtime, Context Policy, Session Replay, Trace, Memory, Hosted Harness.

This series is better for distilling design patterns that remain stable across multiple Agent systems:

Context is not messages[]; it is attention governance.
Memory is not a cache; it is knowledge with scope and provenance.
Tool calls are not text; they are events with permissions, causality, and auditability.
Compression is not just summarization; it is traceable semantic distillation.
A Subagent is not just launching another model; it is a mechanism for context isolation and responsibility reclamation.

The way to read this series is simple: each article revolves around one engineering abstraction, first asking why it emerged, then unpacking what it solves, what boundaries it leaves behind, and how it maps onto data models and invariants.

Article Path

  1. The New Context Manager Paradigm: An Agent’s Attention Operating System
    Upgrades context from stitching together messages[] into an attention governance layer where event facts, state projections, long-term memory, external evidence, tool environments, and system policies all participate.

  2. Agent Long-Term Memory and Self-Optimization Paradigm
    Turns an Agent’s experience into traceable memories, reusable skills, and testable optimizations, while using evaluation, gradual rollout, and rollback mechanisms to help the Agent improve steadily.

  3. The New Tool Manager Paradigm: An Agent’s Action Operating System
    Upgrades the tool system from a function registry into an action governance layer centered on capabilities, intent, permissions, credentials, sandboxes, execution, side effects, and auditing.

  4. Agent Team Task Assignment and Communication Protocols
    Moves an Agent Team beyond a topology of multiple agents into a collaborative runtime for task assignment, leases, structured communication, evidence flowback, conflict handling, and verified merging.

This column will continue to cover topics such as memory governance, collaboration isolation, permission strategies, and auditable execution.

GitHub: 00-series-guide.md