Agentic AI from prototype to production: our EARL 2026 talk

On 8 October 2026, Seth Shenbanjo, our Managing Director, spoke at EARL 2026 in Brighton. Datacove runs EARL. It is a conference for data science and analytics teams that use R and Python.

The title of the talk was "From Prototype to Production: Building Agentic AI Systems That Actually Work in the Enterprise". The talk was for data scientists, analytics leads and engineering managers. These people saw many agent demonstrations. Now they must put an agent into production, and their organisation must trust it.

To get a copy of the slides, contact us.

Usually, the model is not the problem

An agent is a language model that selects and uses tools to do a task with many steps. A demonstration shows that the model works. Production tests all the other parts of the system:

• Who owns each decision?
• Which data can the agent see?
• How does the agent connect to older systems?
• Can a person explain an answer some months later?

We started the talk with six frequent causes of failure for agents in production:

• No person owns the decision.
• The answer has no source that a person can examine.
• The search returns data that the user must not see.
• An integration stops without an alert when a system changes.
• No person can replay the steps that gave an answer.
• The quality of the answers becomes worse after a change to the prompt or to the model.

Five patterns that work in production

Each cause of failure has one pattern that prevents it. The slides show a diagram for each pattern. One slide also shows a short Python example.

1. Design the workflow first. Then design the agent.
2. If the agent has no permitted evidence, it gives no output.
3. Put each system behind a typed tool that has permissions.
4. Record enough data to replay each decision.
5. Test the agent as you test software. Monitor the agent as you monitor a process.

Other subjects in the slides

• Lessons from Eclipse, our Python framework for AI systems. These lessons include why we put guardrails in the platform and not in the prompt.
• How LLM-enabled business intelligence goes past single prompts to answers that people can examine.
• Anonymised examples from professional services, the public sector, financial services and health.
• A score with five questions. Use it with your team to find how ready you are for agentic AI.

Get the slides

To get the slides, use our contact form. Write "EARL 2026 slides" in your message. We will send the slides to you. If you attended EARL 2026, Datacove also shares the slides with attendees.

If your team also wants to move an agent from prototype to production, read about our applied AI engineering work.

Thank you to Datacove for EARL 2026. Thank you also to all the people who came to Session Four.