# What separates an agent that survives production from a demo.

> My view on what separates an agent that survives production from a demo: evals from real failures, scoped tools, replayable traces, and one named owner.

Canonical: https://alexcinovoj.com/services/custom-ai-agents

Almost every agent demos well. The ones that last share four traits, and none of them are about the model.

Drawn from production AI systems Alex has shipped, and the ones that broke on the way there.

## Plan, act, recover. Not autocomplete with a logo.

### Planner

Decomposes a request into steps. Reasons over your domain, your tools, and your data.

### Tool layer

MCP servers, REST clients, database adapters. Tools have schemas, timeouts, and audit logs.

### Memory

Short-term scratch, long-term in Postgres with RLS. No leaking state across users.

### Eval harness

Golden tasks, regressions, model upgrade guards. You see quality before customers do.

## Claude. MCP. Postgres. TypeScript.

Anthropic Claude across API, Bedrock, and Vertex. MCP for tool surfaces. Postgres with Row Level Security for state and permissions. TypeScript end to end. n8n when orchestration is the bottleneck. Helicone, OpenTelemetry, or Langfuse for observability.

## Hierarchies only after a single agent works.

ClawKeeper runs 110 agents in a deterministic finance hierarchy. CipherClaw debugs multi-agent systems in production. Both started as single agents that earned the right to spawn more.

## Shipped systems, not slideware.

Completed 9 Anthropic Academy courses (Agent Engineering, Claude Code, MCP, Context Engineering). Lovable Champion and community leader. 12+ live AI products with auth, billing, and customers. Open-source agent tooling on GitHub.

## Common questions.

## Have an agent that almost ships?

Hardening an agent into something your team can own runs through Systems Decision Audit at TechTide AI, where the scope and the pricing live.