Your agent will crash mid-task: A hands-on lab in durable agent orchestration

Most AI agents look great in a demo and fall apart in production. Tool calls hang. A crash wipes the agent’s state halfway through a task. The same input takes a different path every run. And when something goes wrong, there’s no audit trail to tell you why.

Platform teams are starting to inherit this problem. Every app team building an agent ends up reinventing retries, timeouts and logging, usually badly. In this 60-minute lab, you’ll learn a pattern that fixes this once: split the work between a deterministic execution plane that owns state, retries and control flow, and an LLM reasoning plane that only handles judgement and synthesis.

You’ll apply it to a realistic scenario: an incident-triage agent that reads an alert, queries logs and runbooks through MCP tools, proposes a remediation, and waits for on-call approval before anything runs.

Starting from a provided workflow skeleton, you’ll:

• Add retries, timeouts and a fallback branch to the deterministic workflow

• Wire in an LLM agent that calls MCP tools and drafts a remediation plan

• Gate execution behind a human-in-the-loop approval step, so the agent’s judgement drives the workflow without bypassing it