Assemble capability and execution

Question this part answers

How do you build the system around the agent?

Chapters

1. 11

   The Agent Factory ↗

   Create and govern agents, skills, tools, profiles, and configurations.

2. 12

   Skills as packages ↗

   Package and evaluate reusable capabilities rather than loose prompt files.

3. 13

   Control plane, orchestrator, and execution plane ↗

   Separate authority, coordination, and work.

4. 14

   Durable execution: tasks, attempts, leases, and recovery ↗

   Survive failures without losing truth.

5. 15

   Coding harnesses and agent protocols ↗

   Own the execution contract and place interoperability protocols deliberately.

6. 16

   Harness engineering ↗

   Engineer reproducible sessions, loops, controls, events, and conformance evidence.

7. 17

   Development environments, sandboxes, and compute ↗

   Contain side effects and prove cleanup.

8. 18

   Agent architecture: loop, MCP, tools, context, and memory ↗

   Design one bounded worker as a governed runtime composition.

9. 19

   Data, knowledge, and semantic engineering ↗

   Prepare attributable knowledge and semantic contracts.

10. 20

    Context engineering ↗

    Compile the minimum sufficient, permission-aware Context Package.

11. 21

    Models and capability selection ↗

    Treat models as qualified, replaceable capability profiles.

12. 22

    Routing and the escalation ladder ↗

    Route work to the cheapest qualified capability and escalate on evidence.

13. 23

    Agent and loop engineering ↗

    Choose the least agentic mechanism and coordinate through explicit contracts.

14. 24

    Loop engineering patterns and defaults ↗

    Bound retries and cost, detect oscillation, and stop honestly.

15. 25

    The 12-layer production AI agent stack ↗

    Place each production responsibility in a named layer.

16. 26

    Autonomous engineering workflows ↗

    Give each workflow its own trigger, proof shape, authority, and autonomy.

Retain this

• Use the lightest orchestration model that satisfies the workflow.
• Capabilities are versioned supply, not loose prompts.
• The sandbox defines what intelligence is allowed to affect.