Kihagyás

Knowledge Base

Agent Architecture

cl24s/knowledge-base

Agent Architecture¶

Az Agent Architecture azt vizsgálja, hogyan illeszkednek a modellek, skillek, agentic loopok, state, memory, retrieval, toolok, policyk és external systemek egy maintainable production software architecture-be.

A központi mental model:

Business / Product Use Case
          ↓
Application Layer
          ↓
Agent Runtime / Orchestrator
  ├── canonical state
  ├── context builder
  ├── policy / authorization
  ├── skills
  ├── capabilities
  ├── budgets / recovery
  └── observability
          ↓
Application Ports
          ↓
Infrastructure Adapters
  ├── LLM providers
  ├── retrieval/search/vector stores
  ├── MCP/connectors
  ├── databases
  ├── queues
  ├── sandboxes
  └── external APIs

A model probabilisztikus semantic decision component; nem tulajdonolja az applicationt, canonical state-et, authorizationt vagy side-effect guarantee-ket.

Roadmap¶

1. Agent Architecture Mental Model ¶

Model vs agent vs runtime vs application; probabilistic core + deterministic shell; control plane és execution plane; explicit architecture boundaryk.

Status: detailed first draft complete.

2. Application Structure in the AI Era ¶

Modular monolith, Hexagonal Architecture / Ports & Adapters, Clean Architecture, DDD/bounded contexts, service extraction és dependency direction az AI korszakában.

Status: detailed first draft complete.

3. Core Agent Runtime and Orchestrator ¶

Run lifecycle, state machine, orchestrator responsibility, model gateway, capability registry, policy, checkpoints, cancellation és durable execution.

Status: detailed first draft complete.

4. Context Architecture ¶

Context mint canonical state-ből és evidence-ből épített purpose-specific projection; provenance, trust, freshness, budgets és context builder.

Status: detailed first draft complete.

5. Tool and Capability Architecture ¶

Capability vs tool vs port vs adapter; domain/application service reuse, provider normalization, permission, idempotency és MCP mint adapter mechanism.

Status: detailed first draft complete.

6. Skill Architecture ¶

Skill ownership, registry/discovery, typed contract, capability dependency, versioning, permission architecture, context ownership és eval lifecycle.

Status: detailed first draft complete.

7. State and Memory Architecture ¶

Domain state, run state, session state, long-term memory és model context különválasztása; durable state, concurrency, memory policy és retention.

Status: detailed first draft complete.

8. Knowledge and Retrieval Architecture ¶

RAG mint evidence subsystem; ingestion vs serving; normalized document, chunking, hybrid retrieval, provenance, freshness, ACL és retrieval evaluation.

Status: detailed first draft complete.

9. Single Agent, Workflow and Multi-Agent Patterns ¶

Deterministic workflow, agentic islands, single agent, planner/executor, supervisor/workers, handoff, fan-out/fan-in és multi-agent decision criteria.

Status: detailed first draft complete.

10. Security and Trust Architecture ¶

Authentication/authorization, confused deputy, capability scope, prompt injection, approval architecture, sandboxing, tenant isolation, MCP/tool trust és kill switch.

Status: detailed first draft complete.

11. Reliability, Scaling and Production Runtime ¶

Queues, durable run state, workers, at-least-once semantics, retry/idempotency, backpressure, rate limiting, circuit breaker, bulkhead és scaling.

Status: detailed first draft complete.

12. Evaluation, Observability and Architecture Patterns ¶

End-to-end traces, structured events, layered evaluation, trajectory metrics, replay, regression gate, failure injection és production architecture anti-patternök.

Status: detailed first draft complete.

A software architecture fő elvei¶

Az AI application ugyanúgy profitál a klasszikus architecture elvekből:

high cohesion
low coupling
explicit interfaces
clear state ownership
dependency inversion
bounded contexts
testable boundaries
observable execution

Sőt, ezek még fontosabbak, mert a model behavior probabilisztikus és az AI infrastructure gyorsan változik.

Modular monolith vs microservices¶

Defaultként gyakran erős:

Modular Monolith
├── business/domain modules
├── application/use cases
├── agent runtime
└── infrastructure adapters

Service extraction csak valós okból:

independent scaling
security/isolation
separate ownership
availability requirement
special runtime dependency
independent deployment cadence

Az „AI van benne” önmagában nem microservice requirement.

A dependency direction¶

Business / Domain
       ↑
Application Ports
       ↑
Agent Runtime / Skills
       ↑
Infrastructure Adapters
       ↑
LLM / Vector DB / MCP / Cloud / External APIs

Provider detail kifelé menjen; business semantic befelé stabil maradjon.

Agent vs workflow vs multi-agent¶

Hasznos decision order:

Deterministic code
      ↓
Workflow + bounded agentic island
      ↓
Single modular agent
      ↓
Multi-agent only with a concrete reason

A multi-agent nem „advanced mode”; plusz coordination, state, cost, latency, security és observability complexity.

A teljes tanulási ív¶

AI Foundations
      ↓
Agent Skills
      ↓
Agentic Loops
      ↓
Agent Architecture

Foundations: model, context, tools, structured output, reliability és evaluation alapok.
Skills: reusable task-level capability contractok.
Loops: iterative, bounded execution.
Architecture: mindez hogyan lesz normál production software system.

Készültségi állapot¶

Az Agent Architecture 12/12 first-draft fejezete elkészült. A következő AI deep-dive területet külön érdemes választani: RAG, MCP vagy Memory természetes folytatás lehet.