# Reliable Agents in Production

> A workshop format for moving from an impressive agent demo to an operable system with explicit evidence and controls.


Source: https://andreasnissen.dev/talks/reliable-agents-in-production/index.md


## The question

What has to change between a useful agent prototype and a system an enterprise can operate with confidence?

## Audience

Solutions architects, engineering leaders, platform teams, security teams, and technical executives evaluating production agent systems.

## Verified delivery

I helped run an Amazon Bedrock AgentCore Immersion Day for more than 100 participants. As a distinct part of that program, I designed and delivered a custom evaluation and policy module extension for a cohort of about 30 participants.

My ownership claim is deliberately narrow: it covers the evaluation and policy extension, not every module in the larger Immersion Day.

## Discussion path

1. Define the agent's outcome, authority, and failure boundary.
2. Separate context, reasoning, control, and execution.
3. Turn broad capabilities into narrow tool contracts.
4. Design evaluation across answer quality, plans, policies, and actions.
5. Add observable evidence for model calls, tools, approvals, and outcomes.
6. Walk through failure modes and the operating response.

## Participant outcome

Participants leave with a concrete architecture checklist and a clearer distinction between model capability, system reliability, and execution control.

## Public artifacts

The original learning materials are not published here. The same public-safe engineering argument can be inspected through:

- [The Hard Part of Agentic AI Starts After the Demo](/writing/agentic-ai-after-the-demo/)
- [Mistral Reliability Lab](/projects/mistral-playground/)
- [Context and Control Are Different Layers](/writing/context-and-control/)