Everything around the model that turns an AI agent into a product customers can trust.
Loops, instructions, context, tools, verification, permissions, credentials, cost and evals, framed as the decisions a product manager owns, with a ten-worksheet toolkit for the harness PRD.
Guide PDFs and toolkits download with paid Premium while it is active, or with complimentary access we have approved. Days earned through referrals open the rest of the site, not these files.
Decide / verify / contain / measure
Pages
96
Chapters
31
Diagrams
28
Dated sources
36
Worksheets
10
What is harness engineering, and why does a PM own it?
Two teams can use the same model and ship agents that behave nothing alike. The difference is the harness: what the agent is told and shown, which tools it can call, who checks its work, what it is forbidden to do and what each task may cost. For an AI product, that software is most of the product.
This guide follows one fictional support agent that can issue account credits, from its first runtime choice to a security review, a cost spike and three incidents. Every chapter ends in the decisions a PM owns: what done means, which actions need a person, what a task may cost and what evidence a release needs.
THREE COMPLETE PAGES, FREE TO READ
Judge the reasoning before you decide.
Each preview is a real page from the PDF. Open one to read it at full size.
Page 39
Take "done" away from the model
Why agents approve their own work, and the verification contract that cut reopened tickets from 31% to 6%.
Page 49
Deny first, then ask, then allow
A permission table for a support agent that can issue credits, with the reason for every rule.
Page 63
Find out where the money goes
Why input tokens dominate an agent bill, and the cost per verified resolution to price against.
31 CHAPTERS / 5 PARTS
From the product decision to a design you can defend.
The harness is the product
Learn what surrounds the model, read the same-model evidence critically, use one map for every review and decide what to adopt, extend or build.
01Know what you are actually shippingp. 7
02Read the evidence that the harness mattersp. 11
03Use one map for every reviewp. 14
04Decide how much harness to write yourselfp. 16
Inside the harness
Specify the loop and its endings, a short brief, a cacheable context, memory, tools, skills, retrieval, sub-agents, verification and sessions.
05Specify the loop and how it stopsp. 20
06Write instructions as a short mapp. 22
07Decide what the model sees, and keep it cacheablep. 25
08Give state an owner and a homep. 27
09Design tools around the actions customers care aboutp. 29
10Package procedures as skillsp. 33
11Treat everything the agent reads as untrustedp. 35
12Add agents only when coordination paysp. 37
13Take "done" away from the modelp. 39
14Design the start and end of every sessionp. 42
Secure the harness
Build the threat model from real incidents, choose permissions, shrink credentials, contain code and traffic, and record enough to replay any session.
15Build the threat model from how agents actually failp. 45
16Choose a permission model you can defendp. 48
17Shrink what any credential can reachp. 51
18Contain where code runs and where data goesp. 54
19Record enough to replay any sessionp. 58
Cost, quality and operations
Find where the money goes, rank cost levers by evidence, set budgets, build the customer eval, separate signal from noise and operate upgrades.
20Find out where the money goesp. 63
21Pull the cost levers in order of evidencep. 66
22Set budgets from data and name every failurep. 68
23Build the eval from the customer's own workp. 71
24Separate a real improvement from noisep. 73
25Let models improve the harness, under reviewp. 75
26Operate the harness like a servicep. 77
Prove and defend
Adapt the harness to a regulated domain, check known anti-patterns, write the harness PRD, work three incidents and defend the design.
27Adapt the harness to a new domainp. 80
28Check the design against known anti-patternsp. 82
29Write the harness PRDp. 84
30Work through three incidentsp. 86
31Defend the design in a review or an interviewp. 88
WORKED INCIDENTS
Three incidents, worked from the evidence.
Each fictional incident asks what a PM has to answer: what is known, which record tells the truth, who owns the next step and what changes so it cannot recur.
Destination from untrusted text
The ticket that asked for a new reset address
Treat it as a takeover, find the other affected accounts and make the reset tool read the verified address itself.
Connector update / cache collapse
The cost spike after a quiet release
Trace the spike through the phase breakdown, roll back the connector and make cache hit rate a release gate.
Better average / worse pass^k
The upgrade that lowered one slice
Route the affected slice to the previous model, fix the skill and add a reliability floor for every slice.
INCLUDED / 14 FILES
Ten worksheets for the harness PRD.
The guide ships with ten editable worksheets and three spreadsheets that become the sections of a harness PRD and the evidence a security, finance or compliance reviewer will ask for.
Start with the ownership map and the verification contract. Together they answer the two questions every reviewer asks first: who decides, and how do you know it worked?
These questions, concepts and lessons pick up where the chapters leave off. Each link also sits, clickable, in the chapter it supports, and keeps its own access level on the site.
Everything around the model that turns its proposals into accountable work: the loop, instructions, context, tools, memory, verification, permissions, credentials, sandbox, budgets and audit. The model proposes; the harness decides what actually happens.
Do I need to be an engineer to use this guide?
No. It is written for product managers and explains each part of the harness through the decisions you own and the questions to ask in a review. It never asks you to write code.
How is this different from prompt engineering?
Prompts are one small input to the harness. A prompt can ask an agent not to issue a large refund; a harness can make it impossible. The guide covers the parts that enforce behavior as well as the parts that encourage it.
Which research and products does it cover?
Same-model studies from Anthropic, OpenAI, LangChain, Cursor and Google Research, the 2026 incident record, and dated documentation for the main agent runtimes, with 36 sources checked in October 2026. Company accounts are labelled as company-reported.
How current is it?
Runtime and product facts are dated October 2026 and kept on one comparison page. The rest teaches patterns that outlast the products, so the guide stays useful when tools change.
What is in the toolkit?
14 files: ten editable worksheets from the ownership map to the incident review, plus CSV versions of the runtime scorecard, the permission matrix and the cost budget. Each has one worked example from the guide.
Will this help in AI PM interviews?
Yes. Chapter 31 has ten design prompts with a scoring rubric, and every chapter links to related practice questions and concepts on the site. Agent design, evals and cost come up in most AI PM loops.
Is it included in Premium?
Yes. The full PDF and toolkit come with paid Premium while it is active, or with complimentary access we have approved, and there is nothing extra to buy. Anyone can read three full sample pages first. Days earned through referrals open the rest of the site, not the guide files.
How was it produced?
AI tools helped with research, source checks, drafting and the cover art. The diagrams are vector figures written in code. Fernhill, Resolve, Kinsale Broadband and every person in its scenarios are invented. This edition is ready for publisher review, and automated or AI review is no substitute for an independent human editor signing it off.
