↓Skip to main content

Agent demos work. Agent production is a different sport.

·371 words·2 mins
Table of Contents

Angle: Deflates the demo hype. The demo star is the agent. The production star is boring infrastructure: durable execution, checkpoints, replay, traces. Controversial because it tells builders that the “cool” part is the easy part.

Template: purple cover + purple slides (Torchenstein brand)

Slide 1 — cover #

Headline: Agent demos work. Production is a different sport.

Slide 2 #

Headline: In the demo, the agent is the star.

Supporting copy: It reasons. It calls tools. It solves the task. The audience applauds.

Slide 3 #

Headline: In production, the agent is 20% of the problem.

Supporting copy: The other 80%:

  • durable execution and checkpoints
  • replay after partial failure
  • structured traces and audit
  • service discovery across teams

Slide 4 #

Headline: Agent systems should be designed as long-running distributed processes.

Supporting copy: If you design your agent with those principles in mind, you are will start think differently, start to solve the right problems, and start to build the right system.

Slide 5,6,7 #

Headline: Before building the next agent, think about this.

Supporting copy: Can your system:

  • Resume after a crash without repeating expensive calls?
  • How fixed is the communication between the agents? Did you introduce any assumptions about the communication between the agents? Queues, endpoints to call, even MCP to communicate between the agents?
  • Agent user authorization and authentication? How you are going to handle it?

Companion post copy (text to accompany the carousel in LinkedIn feed):

Agent demos are dangerously convincing.

The agent reasons, calls tools, solves the task. Everyone applauds. Then you try to ship it.

Suddenly the hard part is not the graph and agent prompts. It is everything around it: → What happens when the agent crashes mid-task? → Can it resume without re-running expensive tool calls? → Communication between the agents? → Can you handle agent user authorization and authentication?

These are not AI problems. They are distributed systems problems. Durable execution, checkpoints, replay, idempotent side effects, structured traces.

The uncomfortable realization: agent graph or chain design is not enough. You need to have a good agentic engineering mindset to build a successful agentic system.

The best agent systems will not win on model quality alone. They will win on infrastructure nobody finds exciting to build.

#AIEngineering #AIAgents #MultiAgentSystems