Skip to content
Building Agents in Production
The Trove library

Building Agents in Production

Build AI agents that survive contact with real users, real data, and real organizations.

Read a preview

You can read the contents and the first 2,000 characters.
Pro includes the whole book.

  • By Sahar Carmel
  • 8 chapters
  • Updated Nov 1, 2026
  • 42 readers

What’s inside

  • The Missing Phase
  • Tools Over RAG
  • The Design Contract
  • The Knowledge Problem
  • The Finite Mind
  • The Org Chart

What is Building Agents in Production?

Build AI agents that survive contact with real users, real data, and real organizations. Drawn from building Trove's own agent team and from consulting work with a large consumer company, it covers why most agent rollouts fail before the code ships, why giving agents tools beats retrieval search, how to design an agent by watching the human expert do the job first, and how to manage memory, orchestration and observability once agents run in production. Each chapter ends with rules and warning signs you can apply to your own system.

Read more

markdown · Listed Sep 1, 2026

  • agents
  • production
  • multi-agent
  • orchestration
  • claude-code

From the first page

Who it's for: Anyone building or advising on agents in production — inside a company, for clients, or as a product. The lessons are framework-agnostic; the examples come from real systems — chiefly the agent team that runs CandleKeep — and generalize to anyone building on top of an agent SDK.

What it is: Everything I learned building agents in production, distilled into a practitioner's guide. Every chapter ends with Rules an agent can apply and Smells you can use to diagnose your own system.

I built two CEOs a personal assistant. Neither of them ever used it.

I want to start here, with a small failure, because it took me a year to understand that it wasn't small at all — that it was, in fact, the most important thing I would learn about building agents, and that almost everyone in this field is about to learn it the hard way.

The two CEOs ran their company across time zones. You could watch the friction in real time: the double-booked calls, the meetings shuffled at midnight, the quiet tax of coordinating lives that spanned too many clocks. So I did the obvious thing. I built them an agent that could correlate their calendars and manage the chaos. And because I am not reckless, I asked them first — would this help? — and they both said what anyone in pain says. Yes. Obviously. Build it.

I built it. I deployed it. I checked in. And it sat there, untouched, like a gym membership bought in January.

Here is the lesson, and it is the foundation of this entire book, so I will state it plainly:

Preview · 2,000 of 161,043 characters

Claude Mods: Building Function-Hook Plugins for Claude Code
Inside Claude Code: The Architecture
The Complete Guide to Building Skills for Claude
Hermes Agent: The Complete Developer Guide

Contents

  1. 01Building Agents in Productionp.1
    1. About This Bookp.1
  2. 02The Missing Phasep.2
    1. The Demo That Killed a Good Agentp.2
    2. The Roadmap Everyone Drawsp.2
    3. What I Teach Before I Build Anythingp.2
    4. The Three Phasesp.2
    5. Lead With the Most Painful Task, Not the Most Impressive Onep.2
    6. Teaching Is Not Only the Workshopp.2
    7. The Press Was Ready. The Readers Were Not.p.2
    8. The One Question That Tells You the Truthp.2
    9. +2 more sections
  3. 03Tools Over RAGp.3
    1. I Learned This Playing Dungeons & Dragonsp.3
    2. Intelligence Is Not the Bottleneckp.3
    3. The Crutch We Don't Need Anymorep.3
    4. Humans Don't Retrieve by Similarityp.3
    5. Give the Agent the Human's Toolsp.3
    6. Give the Agent a Library Card, Not an Indexp.3
    7. A Tool Built for a Human Is Not a Tool Built for an Agentp.3
    8. When RAG Is Still Finep.3
    9. +2 more sections
  4. 04The Design Contractp.4
    1. "That Table Is Garbage"p.4
    2. We Design Agents Backwardsp.4
    3. Watch the Human Firstp.4
    4. Every Failure Is a Blueprintp.4
    5. Freedom to Reason, Constraint to Actp.4
    6. Constrain With Tools, Not Promptsp.4
    7. Domain Artifacts Are Goldp.4
    8. Give It the Bare Metal, or Give It an Orchestratorp.4
    9. +3 more sections
  5. 05The Knowledge Problemp.5
    1. The Competition Is the Tool Itselfp.5
    2. Don't Replace the Driver. Reproduce Him.p.5
    3. The Knowledge Lives in Heads, Not on a Hard Diskp.5
    4. Lean Agentsp.5
    5. Meet the Agent Where the Work Already Livesp.5
    6. What the Loop Taught Mep.5
    7. Rulesp.5
    8. Smellsp.5
  6. 06The Finite Mindp.6
    1. The Context Window Is RAM, Not Diskp.6
    2. Sub-Agents: Borrowing a Fresh Mindp.6
    3. The Agent That Never Remembersp.6
    4. Anatomy of a Memoryp.6
    5. The Index Is the Bottleneckp.6
    6. You Don't Debug Mid-Flightp.6
    7. Memory Rotsp.6
    8. Two Loops, One Trapp.6
    9. +3 more sections
  7. 07The Org Chartp.7
    1. A Sub-Agent Is a Tool. An Agent Is a Worker.p.7
    2. We Are Re-Deriving the Firmp.7
    3. Specialization Is the Power. Coordination Is the Catch.p.7
    4. Orchestration > Parallelismp.7
    5. The Visibility Principlep.7
    6. Governance: You Don't Ask Agents to Behave. You Make Misbehavior Impossible.p.7
    7. Agents Teach Each Other — and the Durability Trapp.7
    8. Three Layers of Soulp.7
    9. +3 more sections
  8. 08The Glass Boxp.8
    1. You Are Running Operations Nowp.8
    2. When It Breaks: Give It a Black Boxp.8
    3. Three Places to Lookp.8
    4. It's Almost Never the Modelp.8
    5. When It Drifts: It Doesn't Show Up in a Log Tailp.8
    6. Hire It, Then Review Itp.8
    7. The HR Agentp.8
    8. What the Review Can't Bep.8
    9. +3 more sections