skills.rest
leonardoacosta
leonardoacostaCOMMUNITY
Explore all skills by this author →
GitHub

agent-architecture

Designs multi-agent AI systems with orchestrator patterns, agent taxonomies, and context budgets.

Updated Aug 2, 2026
Search Tags:#context management#multi-agent#orchestration#coordination#tool selection#ai design#agent architecture
One-click install
npx skills add https://github.com/leonardoacosta/agents --skill agent-architecture-leonardoacosta
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: agent-architecture
Source: https://github.com/leonardoacosta/agents/tree/main/skills/agent-architecture
Command: npx skills add https://github.com/leonardoacosta/agents --skill agent-architecture-leonardoacosta

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Designing multi-agent AI systems often fails due to unclear agent roles, unplanned context budgets, and ad-hoc coordination. This Skill provides structured patterns for orchestrator design, agent composition, tool assignment, and coordination so agent systems are reliable and cost-aware from the start. ## Core Features & Use Cases - Orchestration Patterns: Six reusable patterns including sequential pipelines, parallel fan-out/fan-in, iterative loops, hierarchical delegation, gate-sequenced batches, and tiered model review. - Agent Taxonomy and Tool Selection: A naming convention ({domain}-{role}), role-based capability matrix (analyst, architect, engineer, reviewer), and a least-privilege tool assignment decision matrix. - Context and Cost Management: Context budget planning templates, context strategies (progressive disclosure, summary handoff, shared filesystem), token estimation tables, and cost optimization tactics. - Use Case: When planning a multi-project audit system, use this Skill to define a hierarchical orchestration with per-project sub-orchestrators, assign read-only analyst agents for discovery, and set gate checks between parallel implementation batches. ## Quick Start Ask the AI to design an orchestration pattern for a multi-project audit system with parallel agents and validation gates.

Frequently Asked Questions about agent-architecture

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I design a multi-agent AI system architecture?▼

Start with an orchestrator-first separation where the orchestrator routes and synthesizes but never writes code. Then define agents using the {domain}-{role} naming convention, assign tools by least privilege, and choose an orchestration pattern such as sequential pipeline or parallel fan-out.

What orchestration pattern should I use for parallel agent tasks?▼

Use the parallel fan-out/fan-in pattern when tasks are independent and can run simultaneously, such as codebase research plus web research. The orchestrator spawns agents in parallel and synthesizes their potentially conflicting findings afterward.

How do I assign tools to AI agents following least privilege?▼

Give agents only the tools their role requires: analysts and architects get read-only tools (Read, Grep, Glob), engineers get Write, Edit, and Bash, and reviewers get read-only Bash. The orchestrator avoids Bash except for read-only status commands.

How do I manage context limits across multiple agent sessions?▼

Plan a token budget per agent before spawning, allocating space for system prompts, project context, task specs, and working memory. Use strategies like progressive disclosure, summary handoffs between pipeline stages, and context resets when sessions approach 75 percent capacity.

Can AI agents communicate directly with each other?▼

No, direct agent-to-agent communication is not supported in CLI-based agent systems and is unreliable. All communication should be mediated by the orchestrator or passed through shared filesystem outputs that each agent writes to its own path.

When should I avoid using an iterative loop agent pattern?▼

Avoid iterative loops when tasks lack a clear completion check, since agents can get stuck cycling. Always set a maximum iteration count, and prefer gate-sequenced batches when work has natural validation checkpoints between phases.

Related Skills

Explore More Software Engineering Skills→
AesopScottAesopScott
1

multi-agent-capabilities

Document multi-agent system capabilities, tool ownership, permissions, and operational limits.

Community
Advanced
arthrodarthrod

conejo-code

Execute multi-stage coding workflows with adversarial model review and TDD.

Community
Advanced
CODE-SAURABHCODE-SAURABH
2

ai-engineer

Convert unreliable AI prototypes into production-ready systems with validation and observability.

Community
Advanced
QunnnnQunnnn

dispatching-parallel-agents

Coordinate parallel delegation of independent tasks across specialized agents.

Community
Intermediate