Skip to content

Part IV. Tools and Execution

Up to this point, we already have three important layers for the same support agent:

  • platform architecture;
  • the security perimeter;
  • memory and retrieval discipline.

Now it is time to move to the place where an agent stops being "smart text" and starts doing real work: checking status, opening a ticket, involving a human, or stopping safely.

Cover for the tools and execution part

This is exactly where the most expensive mistakes usually appear:

  • incorrect tool calls;
  • unexpected side effects;
  • unstable integrations;
  • repeated actions without idempotency;
  • overly loose access to external systems.

What This Part Solves

In this part, we break down how to design the execution layer so the agent does not talk to the world directly, but works through clear contracts, limits, and safe gateways.

Short path through this part

If you want a fast pass, read it this way:

  • Chapter 8: understand how the runtime chooses the next step and capability;
  • Chapter 9: see how that capability passes through a sandbox and a contract layer;
  • Chapter 10: lock down how execution survives retries, timeouts, and uncertain side effects.

Together, these three chapters give you an execution model you can discuss as a production layer, not as “the agent can call tools”.

In This Part

Where It Leads Next

The next move after this part is Part V: how to observe this execution model in production, define SLOs around it, and stop regressions from coming back through rollout.