Building Production-Grade AI Agents

1. 
   016 min read

   An Agent Is a System, Not a Prompt

   The useful mental model for agentic software starts with an execution loop, not a clever system message.

   Read chapter
2. 
   025 min read

   Tools: Where Reasoning Meets the Real World

   Tool design is not plumbing. It is one of the most important safety and product-design surfaces in an agent system.

   Read chapter
3. 
   038 min read

   MCP: A Shared Capability Layer for Agents

   Model Context Protocol is most useful when capabilities need to be reusable across agents, clients, and frameworks.

   Read chapter
4. 
   046 min read

   Context Engineering: Give the Model Less, but Better

   The best context window is not the biggest one. It is the smallest trustworthy working set needed for the next decision.

   Read chapter
5. 
   055 min read

   RAG vs Live Tools: Knowledge Is Not State

   Use retrieval for unstructured knowledge and live tools for current authoritative state—and know when the same source needs both.

   Read chapter
6. 
   065 min read

   Deterministic vs Probabilistic Orchestration

   The most important architecture decision in an agent system is deciding which parts should not be agentic at all.

   Read chapter
7. 
   076 min read

   Reliability: Idempotency, Retries, and Partial Failure

   Agentic systems inherit every distributed-systems failure mode—plus a probabilistic decision-maker on top.

   Read chapter
8. 
   086 min read

   Security: Identity, ACLs, Prompt Injection, and Blast Radius

   The model is not a trusted principal. Design as though untrusted content can influence it—and make that influence powerless.

   Read chapter
9. 
   094 min read

   Human Confirmation Without Confirmation Fatigue

   User approval is useful, but it is not authorization—and asking “Are you sure?” before everything is not a product strategy.

   Read chapter
10. 
    106 min read

    Scaling Agent Systems: Backpressure, Quotas, and Workload Isolation

    Ten thousand users is not a capacity number. Peak concurrency, latency SLOs, downstream limits, and workload mix are.

    Read chapter
11. 
    115 min read

    Tool Sprawl, Routers, and Specialist Agents

    When an agent has forty tools, the answer is rarely “write a better prompt and hope.” Reduce the effective action space.

    Read chapter
12. 
    126 min read

    Evals: The Test Harness Around a Stochastic Application

    A production agent needs more than unit tests. It needs a repeatable way to measure outcomes, trajectories, safety, cost, and regressions.

    Read chapter
13. 
    135 min read

    Ablation, Variance, and Adaptive Sampling

    Agent evaluation becomes much clearer once you borrow a small amount of statistical vocabulary—without turning product development into a statistics thesis.

    Read chapter
14. 
    146 min read

    Production Observability: From Traces to Better Agents

    The first month after launch should produce more than dashboards. It should produce a better eval suite and a sharper failure taxonomy.

    Read chapter
15. 
    157 min read

    The Production Agent Vocabulary Sheet

    A compact glossary for architecture discussions, design reviews, and building shared language across teams.

    Read chapter

All writing on PFP Labs is personal. The views expressed in this series are my own and do not represent those of my employer or of any vendor mentioned. Nothing published here is based on non-public information, and everything described comes from publicly available documentation and personal projects.