Design and build autonomous AI agents that take actions in your systems — Turkish-first.
14 min readMap the agent loop: Perceive → Reason → Act → Observe. Define tool boundaries explicitly. An agent with too many tools hallucinates tool calls.
Each tool: single responsibility, typed input/output, idempotent where possible. Write tool descriptions in Turkish if the agent's primary language is Turkish — it reduces misuse.
Short-term: conversation window. Long-term: vector store (pgvector). Episodic: structured DB. For Turkish agents: store summaries in Turkish to avoid translation overhead.
Always have a human-in-the-loop for irreversible actions (payments, deletions, emails). Use a confirmation step with a 5-second cancellation window.
Log every tool call with input/output/latency. Trace the full agent trajectory per session. Alert on >3 consecutive tool failures — the agent is likely stuck.