Agent nedir, chatbot'tan farkı?
Chatbot kullanıcı sorusunu cevaplar; agent bir hedef için autonomous adımlar atar. Agent multi-step planning yapar, tool'lar kullanır, ara sonuçlardan öğrenir, gerekirse strateji değiştirir. AIOR projelerinde agent uygulamaları daha kompleks senaryolar için: müşteri onboarding, kompleks support workflow, multi-system integration.Basic agent loop
Tipik agent çalışma döngüsü:- 1. Goal received.
- 2. Plan generated — hangi adımlar gerekli?
- 3. Tool call — bir adım çalıştır.
- 4. Observe result.
- 5. Decide — devam, replan, veya terminate?
- 6. Repeat 3-5 until done or max iterations.
ReAct (Reasoning + Acting) pattern bu loop'un yaygın implementasyonu.
Planning approach'ler
Agent nasıl plan yapar?- Reactive — her step'te bir sonraki action'ı karar verir.
- Plan-then-execute — başta tam plan üretir, sonra adımları çalıştırır.
- Hierarchical — yüksek seviye plan + alt-seviye agent'lar her step için.
- Iterative refinement — initial plan ile başlar, sonuçlara göre revize eder.
AIOR projelerinde iterative refinement standart — basit yeterli ve adaptif.
Tool design — agent için
Agent tool'ları chatbot tool'larından farklı:- Daha "verbose" description — agent ne zaman kullanacağını anlamalı.
- Idempotency — agent retry yapabilir, side effect kontrolü kritik.
- Detailed return — sadece success/failure değil, ne oldu açık şekilde.
- Error context — agent recovery için yeterli bilgi.
Memory ve state management
Multi-step agent'lar memory'e ihtiyaç duyar:- Short-term — current goal'un step'leri ve sonuçları (conversation history).
- Long-term — geçmiş hedefler, öğrenilen pattern'ler (kalıcı DB).
- Working memory — geçici hesaplama sonuçları.
AIOR'da agent state'i Redis'te tutuyoruz — fast access, TTL ile cleanup.
Iteration limit ve safety
Agent sonsuza kadar loop'a girmemeli:- Max iteration count (örn. 20 step).
- Token budget per session.
- Time limit per goal.
- Cost limit per session.
- Circuit breaker — N hata sonrası dur.
Human-in-the-loop
Destructive veya risky action'lar için onay almak şart. AIOR projelerinde:- DELETE veya UPDATE action'larında confirmation step.
- External email gönderme öncesi review.
- Payment veya financial transaction'da explicit approval.
- Uncertain decision'larda human review request.
Observability — agent traces
Agent'ın ne yaptığını izlemek kritik. AIOR'da her agent execution için trace:- Initial goal.
- Generated plan.
- Each step's reasoning.
- Tool calls (name, params, result).
- Time and token cost per step.
- Final outcome.
LangSmith veya custom tracing infrastructure ile bu data izleniyor.
Multi-agent systems
Karmaşık problemler için multiple specialized agent. AIOR'da gördüğümüz pattern'ler:- Researcher agent + writer agent — content production.
- Planner agent + executor agent — task orchestration.
- Critic agent + worker agent — quality control.
Coordination overhead var; ufak problem için tek agent yeterli.
Frameworks — LangChain, LlamaIndex, custom
- LangChain / LangGraph — geniş ekosistem, complex agent için.
- LlamaIndex — RAG-focused, agent ile birleşik.
- CrewAI — multi-agent orchestration.
- Custom code — basit agent'lar için framework gereksiz, custom kod tercih.
AIOR'da framework lock-in'den kaçınıyoruz — custom code + direct LLM API call genelde yeterli.
Eval — agent başarısı nasıl ölçülür?
Agent eval chatbot'tan zor:- Task completion rate.
- Step count efficiency.
- Cost per completed task.
- Tool call accuracy.
- Human intervention rate.
- Quality of final output (LLM-as-judge).
Production deployment
Agent production'da farklı kaygılar:- Background job processing — long-running agent'lar async.
- Queue-based — agent task'leri kuyruğa, worker process'leri.
- Idempotency — restart safe.
- Audit log — compliance için.
- Cost monitoring — runaway prevention.
Sonuç
Agent uygulamaları 2026'da emergent ama olgunlaşan bir alan. Doğru planning, sıkı tool design, human-in-the-loop guardrails ve disiplinli observability ile production-grade agent'lar mümkün. AIOR olarak müşteri agent projelerinde bu pattern'leri standart paket olarak teslim ediyoruz. Sizin agent projenizde en zorlandığınız konu ne — planning quality, tool reliability, yoksa cost control mı?What is an agent? Difference from chatbot?
A chatbot answers a user question; an agent takes autonomous steps toward a goal. Agents do multi-step planning, use tools, learn from intermediate results, and change strategy if needed. Agent applications on AIOR projects fit more complex scenarios: customer onboarding, complex support workflows, multi-system integration.Basic agent loop
Typical agent loop:- 1. Goal received.
- 2. Plan generated — which steps are required?
- 3. Tool call — execute a step.
- 4. Observe result.
- 5. Decide — continue, replan, or terminate?
- 6. Repeat 3-5 until done or max iterations.
ReAct (Reasoning + Acting) is a common implementation of this loop.
Planning approaches
How does an agent plan?- Reactive — decide the next action at each step.
- Plan-then-execute — produce a full plan up front, then execute.
- Hierarchical — high-level plan + low-level agents per step.
- Iterative refinement — start with an initial plan, revise based on results.
Iterative refinement is the standard on AIOR projects — simple, sufficient, and adaptive.
Tool design — for agents
Agent tools differ from chatbot tools:- More verbose descriptions — agent must understand when to use it.
- Idempotency — agents retry; side-effect control is critical.
- Detailed return — not just success/failure but what happened, clearly.
- Error context — enough info for agent recovery.
Memory and state management
Multi-step agents need memory:- Short-term — steps and results of the current goal (conversation history).
- Long-term — past goals, learned patterns (persistent DB).
- Working memory — temporary computation results.
AIOR keeps agent state in Redis — fast access, TTL cleanup.
Iteration limits and safety
Agents must not loop forever:- Max iteration count (e.g. 20 steps).
- Token budget per session.
- Time limit per goal.
- Cost limit per session.
- Circuit breaker — stop after N errors.
Human-in-the-loop
Destructive or risky actions need approval. On AIOR projects:- Confirmation step on DELETE or UPDATE actions.
- Review before external email send.
- Explicit approval on payment or financial transactions.
- Human review request on uncertain decisions.
Observability — agent traces
Tracking what an agent does is critical. AIOR captures a trace per agent execution:- Initial goal.
- Generated plan.
- Each step's reasoning.
- Tool calls (name, params, result).
- Time and token cost per step.
- Final outcome.
Tracked via LangSmith or custom tracing infrastructure.
Multi-agent systems
Multiple specialised agents for complex problems. Patterns we see at AIOR:- Researcher agent + writer agent — content production.
- Planner agent + executor agent — task orchestration.
- Critic agent + worker agent — quality control.
Coordination overhead exists; for small problems a single agent is enough.
Frameworks — LangChain, LlamaIndex, custom
- LangChain / LangGraph — broad ecosystem, for complex agents.
- LlamaIndex — RAG-focused, combined with agents.
- CrewAI — multi-agent orchestration.
- Custom code — for simple agents, frameworks are overkill; custom code preferred.
AIOR avoids framework lock-in — custom code + direct LLM API calls usually suffices.
Eval — measuring agent success
Agent eval is harder than chatbot eval:- Task completion rate.
- Step count efficiency.
- Cost per completed task.
- Tool call accuracy.
- Human intervention rate.
- Quality of final output (LLM-as-judge).
Production deployment
Agent production has different concerns:- Background job processing — long-running agents go async.
- Queue-based — agent tasks in queue, worker processes.
- Idempotency — restart-safe.
- Audit log — for compliance.
- Cost monitoring — runaway prevention.
