How an agent acts and keeps running, in thirteen slides: tool use, a tool definition and its JSON schema, the agent loop and its limits, workflows and agents, five workflow patterns, orchestration as a LangGraph graph, a person approving a write, durable execution through a crash, Temporal workflows and activities, idempotent tool calls, agent runs as DevOps workloads, and MCP.
The slides
The tool is lesson 5 of
LLMOps for AI Agents, in thirteen slides. The on-call
assistant reads metrics, pods and logs with three read-only tools and opens a ticket with one write tool, and these
slides follow how it calls them, how its runs are orchestrated, and how a run survives a crash:
- What is an AI agent?: one agent run with no limits and no saved progress, and the same run with them.
- Tool use: the model asks, our code runs: one question, as a sequence of messages.
- A tool definition: a name, a description and a JSON schema:
get_metrics, the model's call and our result. - The agent loop and its limits: the loop in code, its turn, budget and timeout limits, and one run.
- Workflows and agents: a fixed path in our code beside a path the model chooses.
- Five workflow patterns: chaining, routing, parallelisation, orchestrator-workers and evaluator-optimizer.
- Orchestration as a graph: LangGraph: the assistant's nodes, edges, state and checkpoints.
- Human in the loop: approval before a write: a run paused for hours at
open_ticket, and resumed. - Durable execution: surviving a crash mid-run: a run restarted from scratch beside one replayed from its history.
- Temporal: workflows and activities: the run as a Temporal workflow, with retries and a signal.
- Idempotent tool calls: opening the ticket once: a retry with and without an idempotency key.
- Agent runs as DevOps workloads: a queue, workers, and the limits on every run.
- MCP: tools served by their own servers: three MCP servers, and one
tools/call.
The buttons on the left and right edges move between slides. So do the arrow keys and a swipe.
The site has more on the same subject:
The lab
The questions, the tool outputs, the costs and times of the run, the ticket numbers, the run and call ids and the
example limits are made up for the drawings. The code is a sketch: the loop's model.create and run_tool stand for
our own wrappers, and the LangGraph code leaves out its imports and the building of the graph.
Sources