Glossary
Short definitions of the terms that come up in AI-native software delivery. Every entry links to the guide or article that covers it in full.
AI agent
Software that can take a sequence of actions toward a goal with limited step-by-step human direction, as opposed to a single AI call that answers one prompt and stops. Moving an agent from a working demo to production usually means adding safety checks the demo never needed.
AI-native deliveryAI-native delivery
A way of building software where AI writes the code by default and a spec plus an eval suite, not a person, verify that every change does what it was supposed to do. It differs from AI-assisted development, where a human still writes most of the code with AI as a helper.
AI-native deliveryEval-driven development
A practice where an automated eval suite, written from the specification before the change is released, is the thing that proves AI-generated code works. It does the job a senior reviewer used to do, but it checks the same rules every time instead of remembering them selectively.
Team and engagementFDSE
Forward deployed software engineer, the title Palantir and OpenAI use for the same role others call FDE. The variant exists to signal that the person writes production software rather than producing architecture or running demos.
Cost and scopeFixed price vs time and materials
The two common ways to pay for software work. Fixed price sets one number for a defined scope and tends to make changes costly. Time and materials pays for the hours worked and adapts more easily, but depends on trusting the team doing the work.
Team and engagementForward deployed engineer
A software engineer placed inside a customer's organization to build and release production software against that customer's real data and systems, and who stays accountable until the deployment is running. Three tests tell it apart from similar roles: they merge production code, they work in the customer's environment, and they are answerable if it does not run.
AI products and agentsHallucination
A confident-sounding answer from an AI model that is factually wrong or fabricated. Reducing it in a product usually means limiting what the model is allowed to answer from, because telling it to be more careful is not enough on its own.
ComplianceHIPAA
The US federal law governing how healthcare software must protect patient health information, covering who can access it, how it is transmitted, and what happens when a breach occurs. It affects the technical design of the software as well as a company's paperwork.
AI products and agentsHuman in the loop
A design decision to have a person review or approve an AI system's output before it takes effect, usually reserved for the actions where a wrong AI decision would be expensive or hard to reverse.
Cost and scopeMVP (minimum viable product)
The smallest version of a product that can prove the core idea is worth more investment. It is a different, smaller thing from the final product, built to answer one question as cheaply as possible.
ComplianceNAIC Insurance Data Security Model Law
A model law, adopted in some form by more than 25 US states, that requires insurers to run a written information security program, an annual risk assessment, and a breach notification process. It is the main set of compliance rules insurance software has to be built to meet.
Team and engagementOnshore, nearshore, offshore
Three ways to locate a development team relative to your own time zone and geography: in your own country (onshore), a nearby time zone (nearshore), or a distant one (offshore). Each gives a different balance of cost, shared working hours, and ease of communication.
Team and engagementPalantirization
Adopting Palantir's four practices together: engineers embedded with customers, an opinionated integrated platform, sales with close personal attention to customers whose operations depend on the system, and outcome-based multi-year pricing. An a16z partner coined it as a warning, since companies often take on the expensive practices without the reusable platform that made them profitable.
AI products and agentsRAG (retrieval-augmented generation)
A technique where a system looks up relevant information from a document store before asking a model to answer, so the answer can be based on real, current data instead of only what the model learned during training.
Cost and scopeScope creep
The slow, mostly invisible growth of a project's requirements after the budget and timeline were set, usually arriving as many small additions rather than one big decision. It is one of the most common reasons a fixed budget stops matching the actual build.
ComplianceSOC 2
An audit framework that verifies a company's controls over security, availability, and confidentiality of customer data. Financial and healthcare software buyers frequently require a vendor to hold or be working toward a SOC 2 report before they will sign a contract.
Team and engagementSoftware development partner
A team that helps make the plan for a product and stays accountable for whether it still works after launch, as opposed to a vendor that builds a fixed spec and leaves once it is released.
Team and engagementStaff augmentation
Adding engineers to your existing team and process, under your management, rather than handing a whole project to an outside team. It adds people to do more of the work. A partner, by contrast, also helps decide what gets built.
Cost and scopeTechnical debt
Shortcuts taken to release faster now that cost more to fix later, the same way a financial loan costs more the longer it goes unpaid. Not every shortcut is a mistake; the question is whether the team can see the debt and choose when to fix it.
Team and engagementTechnical due diligence
An independent review of a company's codebase, engineering practices, and technical risk, usually done before an investment or acquisition. It answers whether the software can support the plan the business is being valued on.
AI-native deliveryThe AI last mile
The last mile means the final step before production: the gap between an AI-built demo that works once in front of an audience and a product that keeps working with real users, real data, and unusual cases nobody planned for. Finishing that step is usually most of the actual engineering effort in an AI product.
Team and engagementThe productization loop
The mechanism that turns one customer's custom work into a capability every customer gets: a fixed cadence, a named owner who can promote work into the platform, and a stop rule that halts work only ever serving a single customer. Without it, engineers working inside customer companies build up deployments nobody can maintain.
AI-native deliveryThe verification gap
The difference between how fast an AI model produces code and how fast a team can prove that code is correct. Generating code is now fast and proving it correct is still slow, so most of the real engineering work is now spent in this gap.
ComplianceUnauthorized practice of law (UPL)
A legal restriction on giving specific legal advice without a law license, which limits what a piece of legal software can and cannot say to a user. States differ on where the limit is, and some now explicitly exclude clearly-labeled software tools from the definition.