Challenges with Agentic Systems
Move from demo to production. The hard problems appear immediately.
State Across Restarts
Conversation history in process memory works until the process restarts. Durable state requires solving concurrent modification and session lifecycle.
Reliable Tool Execution
External APIs time out. Subprocesses crash. An agent that loses its place after a tool failure is not production software.
Distributed Agent Comms
Point-to-point HTTP is fragile. If either side restarts mid-delegation, the task is lost. This is the common case at scale.
Large Tool Results
Embedding a full PDF or database result in the LLM context is expensive and often impossible. You need an artifact pipeline.
Access Control
Per-user, per-agent, per-tool permissions must propagate when Agent A delegates to Agent B, which calls Tool C on behalf of User D.
Reactive Execution
The highest-value agents are triggered by events, not human chat. Building this on HTTP means building a polling layer your app now has to maintain.